update v0.2.0

.github/workflows/docs.yml

@@ -3,8 +3,16 @@ on:
   push:
     branches:
       - main
+  pull_request:
+    branches:
+      - main
+
 permissions:
   contents: write
+  pull-requests: write
+  deployments: write
+  pages: write
+
 jobs:
   deploy:
     runs-on: ubuntu-latest
@@ -17,12 +25,26 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: 3.x
-      - run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV 
+      - run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV
       - uses: actions/cache@v4
         with:
           key: mkdocs-material-${{ env.cache_id }}
           path: .cache
           restore-keys: |
             mkdocs-material-
-      - run: pip install mkdocs-material 
-      - run: mkdocs gh-deploy --force
+      - run: pip install mkdocs-material
+      - name: Build docs
+        run: mkdocs build
+
+      - name: Deploy to GH Pages (main)
+        if: github.event_name == 'push'
+        run: mkdocs gh-deploy --force
+
+      - name: Deploy PR Preview
+        if: github.event_name == 'pull_request'
+        uses: rossjrw/pr-preview-action@v1
+        with:
+          source-dir: ./site
+          preview-branch: gh-pages
+          umbrella-dir: pr-preview
+          action: auto

.gitignore

@@ -13,6 +13,12 @@ node_modules
 backend/**/templates/
 demo/MobileNetSSD_deploy.caffemodel
 demo/MobileNetSSD_deploy.prototxt.txt
+demo/scratch
+.gradio
+.vscode
 .DS_Store
 test/
-.env
+.env
+
+.vscode/*
+.venv*

CNAME

@@ -0,0 +1 @@
+fastrtc.org

README.md

@@ -1,9 +1,13 @@
-<h1 style='text-align: center; margin-bottom: 1rem'> Gradio WebRTC ⚡️ </h1>
+<div style='text-align: center; margin-bottom: 1rem; display: flex; justify-content: center; align-items: center;'>
+    <h1 style='color: white; margin: 0;'>FastRTC</h1>
+    <img src='https://huggingface.co/datasets/freddyaboulton/bucket/resolve/main/fastrtc_logo_small.png'
+         alt="FastRTC Logo" 
+         style="margin-right: 10px;">
+</div>
 
 <div style="display: flex; flex-direction: row; justify-content: center">
-<img style="display: block; padding-right: 5px; height: 20px;" alt="Static Badge" src="https://img.shields.io/pypi/v/gradio_webrtc"> 
-<a href="https://github.com/freddyaboulton/gradio-webrtc" target="_blank"><img alt="Static Badge" style="display: block; padding-right: 5px; height: 20px;" src="https://img.shields.io/badge/github-white?logo=github&logoColor=black"></a>
-<a href="https://freddyaboulton.github.io/gradio-webrtc/" target="_blank"><img alt="Static Badge" src="https://img.shields.io/badge/Docs-ffcf40"></a>
+<img style="display: block; padding-right: 5px; height: 20px;" alt="Static Badge" src="https://img.shields.io/pypi/v/fastrtc"> 
+<a href="https://github.com/freddyaboulton/fastrtc" target="_blank"><img alt="Static Badge" src="https://img.shields.io/badge/github-white?logo=github&logoColor=black"></a>
 </div>
 <div align="center">
 <strong>中文|<a href="./README_en.md">English</a></strong>
@@ -24,12 +28,12 @@ gradio cc build --no-generate-docs
 ```
 
 ```bash
-pip install dist/gradio_webrtc-0.0.30.dev0-py3-none-any.whl
+pip install dist/fastrtc-0.0.15.dev0-py3-none-any.whl
 ```
 
 ## Docs
 
-https://freddyaboulton.github.io/gradio-webrtc/
+[https://fastrtc.org](https://fastrtc.org)
 
 ## Examples
 

README_EN.md

@@ -0,0 +1,182 @@
+<h1 style='text-align: center; margin-bottom: 1rem'> Gradio WebRTC ⚡️ </h1>
+
+<div style="display: flex; flex-direction: row; justify-content: center">
+<img style="display: block; padding-right: 5px; height: 20px;" alt="Static Badge" src="https://img.shields.io/pypi/v/gradio_webrtc"> 
+<a href="https://github.com/freddyaboulton/gradio-webrtc" target="_blank"><img alt="Static Badge" style="display: block; padding-right: 5px; height: 20px;" src="https://img.shields.io/badge/github-white?logo=github&logoColor=black"></a>
+<a href="https://freddyaboulton.github.io/gradio-webrtc/" target="_blank"><img alt="Static Badge" src="https://img.shields.io/badge/Docs-ffcf40"></a>
+</div>
+<div align="center">
+<strong><a href="./README_en.md">中文</a>|English</strong>
+</div>
+This repository is forked from the original gradio_webrtc repository, primarily adding `video_chat` as an allowed parameter to be enabled by default. This mode is consistent with the behavior of the original `modality="audio-video"` and `mode="send-receive"`, but the UI has been rewritten to include more interactive capabilities (more microphone controls, and the ability to display local video information). The visual presentation is shown below.
+
+If `video_chat` is manually set to `False`, its usage is consistent with the original repository https://freddyaboulton.github.io/gradio-webrtc/
+
+![picture-in-picture](docs/image.png)
+![side-by-side](docs/image2.png)
+
+## Installation
+
+```bash
+gradio cc install
+gradio cc build --no-generate-docs
+```
+
+```bash
+pip install dist/gradio_webrtc-0.0.30.dev0-py3-none-any.whl
+```
+
+## Docs
+
+https://freddyaboulton.github.io/gradio-webrtc/
+
+## Examples
+
+When using it, you need a handler as the entry parameter of the component and implement code similar to the following:
+
+```python
+import asyncio
+import base64
+from io import BytesIO
+
+import gradio as gr
+import numpy as np
+from gradio_webrtc import (
+    AsyncAudioVideoStreamHandler,
+    WebRTC,
+    VideoEmitType,
+    AudioEmitType,
+)
+from PIL import Image
+
+
+def encode_audio(data: np.ndarray) -> dict:
+    """Encode Audio data to send to the server"""
+    return {"mime_type": "audio/pcm", "data": base64.b64encode(data.tobytes()).decode("UTF-8")}
+
+
+def encode_image(data: np.ndarray) -> dict:
+    with BytesIO() as output_bytes:
+        pil_image = Image.fromarray(data)
+        pil_image.save(output_bytes, "JPEG")
+        bytes_data = output_bytes.getvalue()
+    base64_str = str(base64.b64encode(bytes_data), "utf-8")
+    return {"mime_type": "image/jpeg", "data": base64_str}
+
+
+class VideoChatHandler(AsyncAudioVideoStreamHandler):
+    def __init__(
+        self, expected_layout="mono", output_sample_rate=24000, output_frame_size=480
+    ) -> None:
+        super().__init__(
+            expected_layout,
+            output_sample_rate,
+            output_frame_size,
+            input_sample_rate=24000,
+        )
+        self.audio_queue = asyncio.Queue()
+        self.video_queue = asyncio.Queue()
+        self.quit = asyncio.Event()
+        self.session = None
+        self.last_frame_time = 0
+
+    def copy(self) -> "VideoChatHandler":
+        return VideoChatHandler(
+            expected_layout=self.expected_layout,
+            output_sample_rate=self.output_sample_rate,
+            output_frame_size=self.output_frame_size,
+        )
+
+    #Process video data uploaded by the client
+    async def video_receive(self, frame: np.ndarray):
+        newFrame = np.array(frame)
+        newFrame[0:, :, 0] = 255 - newFrame[0:, :, 0]
+        self.video_queue.put_nowait(newFrame)
+
+    #Prepare the video data sent by the server
+    async def video_emit(self) -> VideoEmitType:
+        return await self.video_queue.get()
+
+    #Process audio data uploaded by the client
+    async def receive(self, frame: tuple[int, np.ndarray]) -> None:
+        frame_size, array = frame
+        self.audio_queue.put_nowait(array)
+
+    #Prepare the audio data sent by the server
+    async def emit(self) -> AudioEmitType:
+        if not self.args_set.is_set():
+            await self.wait_for_args()
+        array = await self.audio_queue.get()
+        return (self.output_sample_rate, array)
+
+    def shutdown(self) -> None:
+        self.quit.set()
+        self.connection = None
+        self.args_set.clear()
+        self.quit.clear()
+
+
+
+css = """
+footer {
+	display: none !important;
+}
+"""
+
+with gr.Blocks(css=css) as demo:
+        webrtc = WebRTC(
+            label="Video Chat",
+            modality="audio-video",
+            mode="send-receive",
+            video_chat=True,
+            elem_id="video-source",
+        )
+        webrtc.stream(
+            VideoChatHandler(),
+            inputs=[webrtc],
+            outputs=[webrtc],
+            time_limit=150,
+            concurrency_limit=2,
+        )
+
+
+if __name__ == "__main__":
+    demo.launch()
+
+```
+
+## Deployment
+
+When deploying in a cloud environment (like Hugging Face Spaces, EC2, etc), you need to set up a TURN server to relay the WebRTC traffic.
+The easiest way to do this is to use a service like Twilio.
+
+```python
+from twilio.rest import Client
+import os
+
+account_sid = os.environ.get("TWILIO_ACCOUNT_SID")
+auth_token = os.environ.get("TWILIO_AUTH_TOKEN")
+
+client = Client(account_sid, auth_token)
+
+token = client.tokens.create()
+
+rtc_configuration = {
+    "iceServers": token.ice_servers,
+    "iceTransportPolicy": "relay",
+}
+
+with gr.Blocks() as demo:
+    ...
+    rtc = WebRTC(rtc_configuration=rtc_configuration, ...)
+    ...
+```
+
+## Contributors
+
+[csxh47](https://github.com/xhup)
+[bingochaos](https://github.com/bingochaos)
+[sudowind](https://github.com/sudowind)
+[emililykimura](https://github.com/emililykimura)
+[Tony](https://github.com/raidios)
+[Cheng Gang](https://github.com/lovepope)

backend/fastrtc/__init__.py

@@ -0,0 +1,76 @@
+from .credentials import (
+    get_hf_turn_credentials,
+    get_turn_credentials,
+    get_twilio_turn_credentials,
+)
+from .pause_detection import (
+    ModelOptions,
+    PauseDetectionModel,
+    SileroVadOptions,
+    get_silero_model,
+)
+from .reply_on_pause import AlgoOptions, ReplyOnPause
+from .reply_on_stopwords import ReplyOnStopWords
+from .speech_to_text import MoonshineSTT, get_stt_model
+from .stream import Stream, UIArgs
+from .text_to_speech import KokoroTTSOptions, get_tts_model
+from .tracks import (
+    AsyncAudioVideoStreamHandler,
+    AsyncStreamHandler,
+    AudioEmitType,
+    AudioVideoStreamHandler,
+    StreamHandler,
+    VideoEmitType,
+)
+from .utils import (
+    AdditionalOutputs,
+    Warning,
+    WebRTCError,
+    aggregate_bytes_to_16bit,
+    async_aggregate_bytes_to_16bit,
+    audio_to_bytes,
+    audio_to_file,
+    audio_to_float32,
+    audio_to_int16,
+    wait_for_item,
+)
+from .webrtc import (
+    WebRTC,
+)
+
+__all__ = [
+    "AsyncStreamHandler",
+    "AudioVideoStreamHandler",
+    "AudioEmitType",
+    "AsyncAudioVideoStreamHandler",
+    "AlgoOptions",
+    "AdditionalOutputs",
+    "aggregate_bytes_to_16bit",
+    "async_aggregate_bytes_to_16bit",
+    "audio_to_bytes",
+    "audio_to_file",
+    "audio_to_float32",
+    "audio_to_int16",
+    "get_hf_turn_credentials",
+    "get_twilio_turn_credentials",
+    "get_turn_credentials",
+    "ReplyOnPause",
+    "ReplyOnStopWords",
+    "SileroVadOptions",
+    "get_stt_model",
+    "MoonshineSTT",
+    "StreamHandler",
+    "Stream",
+    "VideoEmitType",
+    "WebRTC",
+    "WebRTCError",
+    "Warning",
+    "get_tts_model",
+    "KokoroTTSOptions",
+    "wait_for_item",
+    "UIArgs",
+    "ModelOptions",
+    "PauseDetectionModel",
+    "get_silero_model",
+    "SileroVadOptions",
+]

backend/fastrtc/credentials.py

@@ -0,0 +1,52 @@
+import os
+from typing import Literal
+
+import requests
+
+
+def get_hf_turn_credentials(token=None):
+    if token is None:
+        token = os.getenv("HF_TOKEN")
+    credentials = requests.get(
+        "https://fastrtc-turn-server-login.hf.space/credentials",
+        headers={"X-HF-Access-Token": token},
+    )
+    if not credentials.status_code == 200:
+        raise ValueError("Failed to get credentials from HF turn server")
+    return {
+        "iceServers": [
+            {
+                "urls": "turn:gradio-turn.com:80",
+                **credentials.json(),
+            },
+        ]
+    }
+
+
+def get_twilio_turn_credentials(twilio_sid=None, twilio_token=None):
+    try:
+        from twilio.rest import Client
+    except ImportError:
+        raise ImportError("Please install twilio with `pip install twilio`")
+
+    if not twilio_sid and not twilio_token:
+        twilio_sid = os.environ.get("TWILIO_ACCOUNT_SID")
+        twilio_token = os.environ.get("TWILIO_AUTH_TOKEN")
+
+    client = Client(twilio_sid, twilio_token)
+
+    token = client.tokens.create()
+
+    return {
+        "iceServers": token.ice_servers,
+        "iceTransportPolicy": "relay",
+    }
+
+
+def get_turn_credentials(method: Literal["hf", "twilio"] = "hf", **kwargs):
+    if method == "hf":
+        return get_hf_turn_credentials(**kwargs)
+    elif method == "twilio":
+        return get_twilio_turn_credentials(**kwargs)
+    else:
+        raise ValueError("Invalid method. Must be 'hf' or 'twilio'")

backend/fastrtc/pause_detection/__init__.py

@@ -0,0 +1,10 @@
+from .protocol import ModelOptions, PauseDetectionModel
+from .silero import SileroVADModel, SileroVadOptions, get_silero_model
+
+__all__ = [
+    "SileroVADModel",
+    "SileroVadOptions",
+    "PauseDetectionModel",
+    "ModelOptions",
+    "get_silero_model",
+]

backend/fastrtc/pause_detection/protocol.py

@@ -0,0 +1,20 @@
+from typing import Any, Protocol, TypeAlias
+
+import numpy as np
+from numpy.typing import NDArray
+
+from ..utils import AudioChunk
+
+ModelOptions: TypeAlias = Any
+
+
+class PauseDetectionModel(Protocol):
+    def vad(
+        self,
+        audio: tuple[int, NDArray[np.int16] | NDArray[np.float32]],
+        options: ModelOptions,
+    ) -> tuple[float, list[AudioChunk]]: ...
+
+    def warmup(
+        self,
+    ) -> None: ...

backend/fastrtc/pause_detection/silero.py

@@ -0,0 +1,329 @@
+import logging
+import warnings
+from dataclasses import dataclass
+from functools import lru_cache
+from typing import List
+
+import click
+import numpy as np
+from huggingface_hub import hf_hub_download
+from numpy.typing import NDArray
+
+from ..utils import AudioChunk
+from .protocol import PauseDetectionModel
+
+logger = logging.getLogger(__name__)
+
+# The code below is adapted from https://github.com/snakers4/silero-vad.
+# The code below is adapted from https://github.com/gpt-omni/mini-omni/blob/main/utils/vad.py
+
+
+@lru_cache
+def get_silero_model() -> PauseDetectionModel:
+    """Returns the VAD model instance and warms it up with dummy data."""
+    # Warm up the model with dummy data
+
+    try:
+        import importlib.util
+
+        mod = importlib.util.find_spec("onnxruntime")
+        if mod is None:
+            raise RuntimeError("Install fastrtc[vad] to use ReplyOnPause")
+    except (ValueError, ModuleNotFoundError):
+        raise RuntimeError("Install fastrtc[vad] to use ReplyOnPause")
+    model = SileroVADModel()
+    print(click.style("INFO", fg="green") + ":\t  Warming up VAD model.")
+    model.warmup()
+    print(click.style("INFO", fg="green") + ":\t  VAD model warmed up.")
+    return model
+
+
+@dataclass
+class SileroVadOptions:
+    """VAD options.
+
+    Attributes:
+      threshold: Speech threshold. Silero VAD outputs speech probabilities for each audio chunk,
+        probabilities ABOVE this value are considered as SPEECH. It is better to tune this
+        parameter for each dataset separately, but "lazy" 0.5 is pretty good for most datasets.
+      min_speech_duration_ms: Final speech chunks shorter min_speech_duration_ms are thrown out.
+      max_speech_duration_s: Maximum duration of speech chunks in seconds. Chunks longer
+        than max_speech_duration_s will be split at the timestamp of the last silence that
+        lasts more than 100ms (if any), to prevent aggressive cutting. Otherwise, they will be
+        split aggressively just before max_speech_duration_s.
+      min_silence_duration_ms: In the end of each speech chunk wait for min_silence_duration_ms
+        before separating it
+      window_size_samples: Audio chunks of window_size_samples size are fed to the silero VAD model.
+        WARNING! Silero VAD models were trained using 512, 1024, 1536 samples for 16000 sample rate.
+        Values other than these may affect model performance!!
+      speech_pad_ms: Final speech chunks are padded by speech_pad_ms each side
+      speech_duration: If the length of the speech is less than this value, a pause will be detected.
+    """
+
+    threshold: float = 0.5
+    min_speech_duration_ms: int = 250
+    max_speech_duration_s: float = float("inf")
+    min_silence_duration_ms: int = 2000
+    window_size_samples: int = 1024
+    speech_pad_ms: int = 400
+
+
+class SileroVADModel:
+    @staticmethod
+    def download_model() -> str:
+        return hf_hub_download(
+            repo_id="freddyaboulton/silero-vad", filename="silero_vad.onnx"
+        )
+
+    def __init__(self):
+        try:
+            import onnxruntime
+        except ImportError as e:
+            raise RuntimeError(
+                "Applying the VAD filter requires the onnxruntime package"
+            ) from e
+
+        path = self.download_model()
+
+        opts = onnxruntime.SessionOptions()
+        opts.inter_op_num_threads = 1
+        opts.intra_op_num_threads = 1
+        opts.log_severity_level = 4
+
+        self.session = onnxruntime.InferenceSession(
+            path,
+            providers=["CPUExecutionProvider"],
+            sess_options=opts,
+        )
+
+    def get_initial_state(self, batch_size: int):
+        h = np.zeros((2, batch_size, 64), dtype=np.float32)
+        c = np.zeros((2, batch_size, 64), dtype=np.float32)
+        return h, c
+
+    @staticmethod
+    def collect_chunks(audio: np.ndarray, chunks: List[AudioChunk]) -> np.ndarray:
+        """Collects and concatenates audio chunks."""
+        if not chunks:
+            return np.array([], dtype=np.float32)
+
+        return np.concatenate(
+            [audio[chunk["start"] : chunk["end"]] for chunk in chunks]
+        )
+
+    def get_speech_timestamps(
+        self,
+        audio: np.ndarray,
+        vad_options: SileroVadOptions,
+        **kwargs,
+    ) -> List[AudioChunk]:
+        """This method is used for splitting long audios into speech chunks using silero VAD.
+
+        Args:
+        audio: One dimensional float array.
+        vad_options: Options for VAD processing.
+        kwargs: VAD options passed as keyword arguments for backward compatibility.
+
+        Returns:
+        List of dicts containing begin and end samples of each speech chunk.
+        """
+
+        threshold = vad_options.threshold
+        min_speech_duration_ms = vad_options.min_speech_duration_ms
+        max_speech_duration_s = vad_options.max_speech_duration_s
+        min_silence_duration_ms = vad_options.min_silence_duration_ms
+        window_size_samples = vad_options.window_size_samples
+        speech_pad_ms = vad_options.speech_pad_ms
+
+        if window_size_samples not in [512, 1024, 1536]:
+            warnings.warn(
+                "Unusual window_size_samples! Supported window_size_samples:\n"
+                " - [512, 1024, 1536] for 16000 sampling_rate"
+            )
+
+        sampling_rate = 16000
+        min_speech_samples = sampling_rate * min_speech_duration_ms / 1000
+        speech_pad_samples = sampling_rate * speech_pad_ms / 1000
+        max_speech_samples = (
+            sampling_rate * max_speech_duration_s
+            - window_size_samples
+            - 2 * speech_pad_samples
+        )
+        min_silence_samples = sampling_rate * min_silence_duration_ms / 1000
+        min_silence_samples_at_max_speech = sampling_rate * 98 / 1000
+
+        audio_length_samples = len(audio)
+
+        state = self.get_initial_state(batch_size=1)
+
+        speech_probs = []
+        for current_start_sample in range(0, audio_length_samples, window_size_samples):
+            chunk = audio[
+                current_start_sample : current_start_sample + window_size_samples
+            ]
+            if len(chunk) < window_size_samples:
+                chunk = np.pad(chunk, (0, int(window_size_samples - len(chunk))))
+            speech_prob, state = self(chunk, state, sampling_rate)
+            speech_probs.append(speech_prob)
+
+        triggered = False
+        speeches = []
+        current_speech = {}
+        neg_threshold = threshold - 0.15
+
+        # to save potential segment end (and tolerate some silence)
+        temp_end = 0
+        # to save potential segment limits in case of maximum segment size reached
+        prev_end = next_start = 0
+
+        for i, speech_prob in enumerate(speech_probs):
+            if (speech_prob >= threshold) and temp_end:
+                temp_end = 0
+                if next_start < prev_end:
+                    next_start = window_size_samples * i
+
+            if (speech_prob >= threshold) and not triggered:
+                triggered = True
+                current_speech["start"] = window_size_samples * i
+                continue
+
+            if (
+                triggered
+                and (window_size_samples * i) - current_speech["start"]
+                > max_speech_samples
+            ):
+                if prev_end:
+                    current_speech["end"] = prev_end
+                    speeches.append(current_speech)
+                    current_speech = {}
+                    # previously reached silence (< neg_thres) and is still not speech (< thres)
+                    if next_start < prev_end:
+                        triggered = False
+                    else:
+                        current_speech["start"] = next_start
+                    prev_end = next_start = temp_end = 0
+                else:
+                    current_speech["end"] = window_size_samples * i
+                    speeches.append(current_speech)
+                    current_speech = {}
+                    prev_end = next_start = temp_end = 0
+                    triggered = False
+                    continue
+
+            if (speech_prob < neg_threshold) and triggered:
+                if not temp_end:
+                    temp_end = window_size_samples * i
+                # condition to avoid cutting in very short silence
+                if (
+                    window_size_samples * i
+                ) - temp_end > min_silence_samples_at_max_speech:
+                    prev_end = temp_end
+                if (window_size_samples * i) - temp_end < min_silence_samples:
+                    continue
+                else:
+                    current_speech["end"] = temp_end
+                    if (
+                        current_speech["end"] - current_speech["start"]
+                    ) > min_speech_samples:
+                        speeches.append(current_speech)
+                    current_speech = {}
+                    prev_end = next_start = temp_end = 0
+                    triggered = False
+                    continue
+
+        if (
+            current_speech
+            and (audio_length_samples - current_speech["start"]) > min_speech_samples
+        ):
+            current_speech["end"] = audio_length_samples
+            speeches.append(current_speech)
+
+        for i, speech in enumerate(speeches):
+            if i == 0:
+                speech["start"] = int(max(0, speech["start"] - speech_pad_samples))
+            if i != len(speeches) - 1:
+                silence_duration = speeches[i + 1]["start"] - speech["end"]
+                if silence_duration < 2 * speech_pad_samples:
+                    speech["end"] += int(silence_duration // 2)
+                    speeches[i + 1]["start"] = int(
+                        max(0, speeches[i + 1]["start"] - silence_duration // 2)
+                    )
+                else:
+                    speech["end"] = int(
+                        min(audio_length_samples, speech["end"] + speech_pad_samples)
+                    )
+                    speeches[i + 1]["start"] = int(
+                        max(0, speeches[i + 1]["start"] - speech_pad_samples)
+                    )
+            else:
+                speech["end"] = int(
+                    min(audio_length_samples, speech["end"] + speech_pad_samples)
+                )
+
+        return speeches
+
+    def warmup(self):
+        for _ in range(10):
+            dummy_audio = np.zeros(102400, dtype=np.float32)
+            self.vad((24000, dummy_audio), None)
+
+    def vad(
+        self,
+        audio: tuple[int, NDArray[np.float32] | NDArray[np.int16]],
+        options: None | SileroVadOptions,
+    ) -> tuple[float, list[AudioChunk]]:
+        sampling_rate, audio_ = audio
+        logger.debug("VAD audio shape input: %s", audio_.shape)
+        try:
+            if audio_.dtype != np.float32:
+                audio_ = audio_.astype(np.float32) / 32768.0
+            sr = 16000
+            if sr != sampling_rate:
+                try:
+                    import librosa  # type: ignore
+                except ImportError as e:
+                    raise RuntimeError(
+                        "Applying the VAD filter requires the librosa if the input sampling rate is not 16000hz"
+                    ) from e
+                audio_ = librosa.resample(audio_, orig_sr=sampling_rate, target_sr=sr)
+
+            if not options:
+                options = SileroVadOptions()
+            speech_chunks = self.get_speech_timestamps(audio_, options)
+            logger.debug("VAD speech chunks: %s", speech_chunks)
+            audio_ = self.collect_chunks(audio_, speech_chunks)
+            logger.debug("VAD audio shape: %s", audio_.shape)
+            duration_after_vad = audio_.shape[0] / sr
+            return duration_after_vad, speech_chunks
+        except Exception as e:
+            import math
+            import traceback
+
+            logger.debug("VAD Exception: %s", str(e))
+            exec = traceback.format_exc()
+            logger.debug("traceback %s", exec)
+            return math.inf, []
+
+    def __call__(self, x, state, sr: int):
+        if len(x.shape) == 1:
+            x = np.expand_dims(x, 0)
+        if len(x.shape) > 2:
+            raise ValueError(
+                f"Too many dimensions for input audio chunk {len(x.shape)}"
+            )
+        if sr / x.shape[1] > 31.25:  # type: ignore
+            raise ValueError("Input audio chunk is too short")
+
+        h, c = state
+
+        ort_inputs = {
+            "input": x,
+            "h": h,
+            "c": c,
+            "sr": np.array(sr, dtype="int64"),
+        }
+
+        out, h, c = self.session.run(None, ort_inputs)
+        state = (h, c)
+
+        return out, state

backend/fastrtc/reply_on_pause.py

@@ -0,0 +1,261 @@
+import asyncio
+import inspect
+from dataclasses import dataclass, field
+from logging import getLogger
+from threading import Event
+from typing import Any, AsyncGenerator, Callable, Generator, Literal, cast
+
+import numpy as np
+from numpy.typing import NDArray
+
+from .pause_detection import ModelOptions, PauseDetectionModel, get_silero_model
+from .tracks import EmitType, StreamHandler
+from .utils import create_message, split_output
+
+logger = getLogger(__name__)
+
+
+@dataclass
+class AlgoOptions:
+    """Algorithm options."""
+
+    audio_chunk_duration: float = 0.6
+    started_talking_threshold: float = 0.2
+    speech_threshold: float = 0.1
+
+
+@dataclass
+class AppState:
+    stream: np.ndarray | None = None
+    sampling_rate: int = 0
+    pause_detected: bool = False
+    started_talking: bool = False
+    responding: bool = False
+    stopped: bool = False
+    buffer: np.ndarray | None = None
+    responded_audio: bool = False
+    interrupted: asyncio.Event = field(default_factory=asyncio.Event)
+
+    def new(self):
+        return AppState()
+
+
+ReplyFnGenerator = (
+    Callable[
+        [tuple[int, NDArray[np.int16]], Any],
+        Generator[EmitType, None, None],
+    ]
+    | Callable[
+        [tuple[int, NDArray[np.int16]]],
+        Generator[EmitType, None, None],
+    ]
+    | Callable[
+        [tuple[int, NDArray[np.int16]]],
+        AsyncGenerator[EmitType, None],
+    ]
+    | Callable[
+        [tuple[int, NDArray[np.int16]], Any],
+        AsyncGenerator[EmitType, None],
+    ]
+)
+
+
+async def iterate(generator: Generator) -> Any:
+    return next(generator)
+
+
+class ReplyOnPause(StreamHandler):
+    def __init__(
+        self,
+        fn: ReplyFnGenerator,
+        startup_fn: Callable | None = None,
+        algo_options: AlgoOptions | None = None,
+        model_options: ModelOptions | None = None,
+        can_interrupt: bool = True,
+        expected_layout: Literal["mono", "stereo"] = "mono",
+        output_sample_rate: int = 24000,
+        output_frame_size: int = 480,
+        input_sample_rate: int = 48000,
+        model: PauseDetectionModel | None = None,
+    ):
+        super().__init__(
+            expected_layout,
+            output_sample_rate,
+            output_frame_size,
+            input_sample_rate=input_sample_rate,
+        )
+        self.can_interrupt = can_interrupt
+        self.expected_layout: Literal["mono", "stereo"] = expected_layout
+        self.output_sample_rate = output_sample_rate
+        self.output_frame_size = output_frame_size
+        self.model = model or get_silero_model()
+        self.fn = fn
+        self.is_async = inspect.isasyncgenfunction(fn)
+        self.event = Event()
+        self.state = AppState()
+        self.generator: (
+            Generator[EmitType, None, None] | AsyncGenerator[EmitType, None] | None
+        ) = None
+        self.model_options = model_options
+        self.algo_options = algo_options or AlgoOptions()
+        self.startup_fn = startup_fn
+
+    @property
+    def _needs_additional_inputs(self) -> bool:
+        return len(inspect.signature(self.fn).parameters) > 1
+
+    def start_up(self):
+        if self.startup_fn:
+            if self._needs_additional_inputs:
+                self.wait_for_args_sync()
+                args = self.latest_args[1:]
+            else:
+                args = ()
+            self.generator = self.startup_fn(*args)
+            self.event.set()
+
+    def copy(self):
+        return ReplyOnPause(
+            self.fn,
+            self.startup_fn,
+            self.algo_options,
+            self.model_options,
+            self.can_interrupt,
+            self.expected_layout,
+            self.output_sample_rate,
+            self.output_frame_size,
+            self.input_sample_rate,
+            self.model,
+        )
+
+    def determine_pause(
+        self, audio: np.ndarray, sampling_rate: int, state: AppState
+    ) -> bool:
+        """Take in the stream, determine if a pause happened"""
+        duration = len(audio) / sampling_rate
+
+        if duration >= self.algo_options.audio_chunk_duration:
+            dur_vad, _ = self.model.vad((sampling_rate, audio), self.model_options)
+            logger.debug("VAD duration: %s", dur_vad)
+            if (
+                dur_vad > self.algo_options.started_talking_threshold
+                and not state.started_talking
+            ):
+                state.started_talking = True
+                logger.debug("Started talking")
+            if state.started_talking:
+                if state.stream is None:
+                    state.stream = audio
+                else:
+                    state.stream = np.concatenate((state.stream, audio))
+            state.buffer = None
+            if dur_vad < self.algo_options.speech_threshold and state.started_talking:
+                return True
+        return False
+
+    def process_audio(self, audio: tuple[int, np.ndarray], state: AppState) -> None:
+        frame_rate, array = audio
+        array = np.squeeze(array)
+        if not state.sampling_rate:
+            state.sampling_rate = frame_rate
+        if state.buffer is None:
+            state.buffer = array
+        else:
+            state.buffer = np.concatenate((state.buffer, array))
+
+        pause_detected = self.determine_pause(
+            state.buffer, state.sampling_rate, self.state
+        )
+        state.pause_detected = pause_detected
+
+    def receive(self, frame: tuple[int, np.ndarray]) -> None:
+        if self.state.responding and not self.can_interrupt:
+            return
+        self.process_audio(frame, self.state)
+        if self.state.pause_detected:
+            self.event.set()
+            if self.can_interrupt and self.state.responding:
+                self._close_generator()
+                self.generator = None
+            if self.can_interrupt:
+                self.clear_queue()
+
+    def _close_generator(self):
+        """Properly close the generator to ensure resources are released."""
+        if self.generator is None:
+            return
+
+        try:
+            if self.is_async:
+                # For async generators, we need to call aclose()
+                if hasattr(self.generator, "aclose"):
+                    asyncio.run_coroutine_threadsafe(
+                        cast(AsyncGenerator[EmitType, None], self.generator).aclose(),
+                        self.loop,
+                    ).result(timeout=1.0)  # Add timeout to prevent blocking
+            else:
+                # For sync generators, we can just exhaust it or close it
+                if hasattr(self.generator, "close"):
+                    cast(Generator[EmitType, None, None], self.generator).close()
+        except Exception as e:
+            logger.debug(f"Error closing generator: {e}")
+
+    def reset(self):
+        super().reset()
+        if self.phone_mode:
+            self.args_set.set()
+        self.generator = None
+        self.event.clear()
+        self.state = AppState()
+
+    async def async_iterate(self, generator) -> EmitType:
+        return await anext(generator)
+
+    def emit(self):
+        if not self.event.is_set():
+            return None
+        else:
+            if not self.generator:
+                self.send_message_sync(create_message("log", "pause_detected"))
+                if self._needs_additional_inputs and not self.args_set.is_set():
+                    if not self.phone_mode:
+                        self.wait_for_args_sync()
+                    else:
+                        self.latest_args = [None]
+                        self.args_set.set()
+                logger.debug("Creating generator")
+                audio = cast(np.ndarray, self.state.stream).reshape(1, -1)
+                if self._needs_additional_inputs:
+                    self.latest_args[0] = (self.state.sampling_rate, audio)
+                    self.generator = self.fn(*self.latest_args)  # type: ignore
+                else:
+                    self.generator = self.fn((self.state.sampling_rate, audio))  # type: ignore
+                logger.debug("Latest args: %s", self.latest_args)
+                self.state = self.state.new()
+            self.state.responding = True
+            try:
+                if self.is_async:
+                    output = asyncio.run_coroutine_threadsafe(
+                        self.async_iterate(self.generator), self.loop
+                    ).result()
+                else:
+                    output = next(self.generator)  # type: ignore
+                audio, additional_outputs = split_output(output)
+                if audio is not None:
+                    self.send_message_sync(create_message("log", "response_starting"))
+                    self.state.responded_audio = True
+                if self.phone_mode:
+                    if additional_outputs:
+                        self.latest_args = [None] + list(additional_outputs.args)
+                return output
+            except (StopIteration, StopAsyncIteration):
+                if not self.state.responded_audio:
+                    self.send_message_sync(create_message("log", "response_starting"))
+                self.reset()
+            except Exception as e:
+                import traceback
+
+                traceback.print_exc()
+                logger.debug("Error in ReplyOnPause: %s", e)
+                self.reset()
+                raise e

backend/fastrtc/reply_on_stopwords.py

@@ -0,0 +1,163 @@
+import asyncio
+import logging
+import re
+from typing import Callable, Literal
+
+import numpy as np
+
+from .reply_on_pause import (
+    AlgoOptions,
+    AppState,
+    ModelOptions,
+    PauseDetectionModel,
+    ReplyFnGenerator,
+    ReplyOnPause,
+)
+from .speech_to_text import get_stt_model, stt_for_chunks
+from .utils import audio_to_float32, create_message
+
+logger = logging.getLogger(__name__)
+
+
+class ReplyOnStopWordsState(AppState):
+    stop_word_detected: bool = False
+    post_stop_word_buffer: np.ndarray | None = None
+    started_talking_pre_stop_word: bool = False
+
+    def new(self):
+        return ReplyOnStopWordsState()
+
+
+class ReplyOnStopWords(ReplyOnPause):
+    def __init__(
+        self,
+        fn: ReplyFnGenerator,
+        stop_words: list[str],
+        startup_fn: Callable | None = None,
+        algo_options: AlgoOptions | None = None,
+        model_options: ModelOptions | None = None,
+        can_interrupt: bool = True,
+        expected_layout: Literal["mono", "stereo"] = "mono",
+        output_sample_rate: int = 24000,
+        output_frame_size: int = 480,
+        input_sample_rate: int = 48000,
+        model: PauseDetectionModel | None = None,
+    ):
+        super().__init__(
+            fn,
+            algo_options=algo_options,
+            startup_fn=startup_fn,
+            model_options=model_options,
+            can_interrupt=can_interrupt,
+            expected_layout=expected_layout,
+            output_sample_rate=output_sample_rate,
+            output_frame_size=output_frame_size,
+            input_sample_rate=input_sample_rate,
+            model=model,
+        )
+        self.stop_words = stop_words
+        self.state = ReplyOnStopWordsState()
+        self.stt_model = get_stt_model("moonshine/base")
+
+    def stop_word_detected(self, text: str) -> bool:
+        for stop_word in self.stop_words:
+            stop_word = stop_word.lower().strip().split(" ")
+            if bool(
+                re.search(
+                    r"\b" + r"\s+".join(map(re.escape, stop_word)) + r"[.,!?]*\b",
+                    text.lower(),
+                )
+            ):
+                logger.debug("Stop word detected: %s", stop_word)
+                return True
+        return False
+
+    async def _send_stopword(
+        self,
+    ):
+        if self.channel:
+            self.channel.send(create_message("stopword", ""))
+            logger.debug("Sent stopword")
+
+    def send_stopword(self):
+        asyncio.run_coroutine_threadsafe(self._send_stopword(), self.loop)
+
+    def determine_pause(  # type: ignore
+        self, audio: np.ndarray, sampling_rate: int, state: ReplyOnStopWordsState
+    ) -> bool:
+        """Take in the stream, determine if a pause happened"""
+        import librosa
+
+        duration = len(audio) / sampling_rate
+
+        if duration >= self.algo_options.audio_chunk_duration:
+            if not state.stop_word_detected:
+                audio_f32 = audio_to_float32((sampling_rate, audio))
+                audio_rs = librosa.resample(
+                    audio_f32, orig_sr=sampling_rate, target_sr=16000
+                )
+                if state.post_stop_word_buffer is None:
+                    state.post_stop_word_buffer = audio_rs
+                else:
+                    state.post_stop_word_buffer = np.concatenate(
+                        (state.post_stop_word_buffer, audio_rs)
+                    )
+                if len(state.post_stop_word_buffer) / 16000 > 2:
+                    state.post_stop_word_buffer = state.post_stop_word_buffer[-32000:]
+                dur_vad, chunks = self.model.vad(
+                    (16000, state.post_stop_word_buffer),
+                    self.model_options,
+                )
+                text = stt_for_chunks(
+                    self.stt_model, (16000, state.post_stop_word_buffer), chunks
+                )
+                logger.debug(f"STT: {text}")
+                state.stop_word_detected = self.stop_word_detected(text)
+                if state.stop_word_detected:
+                    logger.debug("Stop word detected")
+                    self.send_stopword()
+                state.buffer = None
+            else:
+                dur_vad, _ = self.model.vad((sampling_rate, audio), self.model_options)
+                logger.debug("VAD duration: %s", dur_vad)
+                if (
+                    dur_vad > self.algo_options.started_talking_threshold
+                    and not state.started_talking
+                    and state.stop_word_detected
+                ):
+                    state.started_talking = True
+                    logger.debug("Started talking")
+                if state.started_talking:
+                    if state.stream is None:
+                        state.stream = audio
+                    else:
+                        state.stream = np.concatenate((state.stream, audio))
+                state.buffer = None
+                if (
+                    dur_vad < self.algo_options.speech_threshold
+                    and state.started_talking
+                    and state.stop_word_detected
+                ):
+                    return True
+        return False
+
+    def reset(self):
+        super().reset()
+        self.generator = None
+        self.event.clear()
+        self.state = ReplyOnStopWordsState()
+
+    def copy(self):
+        return ReplyOnStopWords(
+            self.fn,
+            self.stop_words,
+            self.startup_fn,
+            self.algo_options,
+            self.model_options,
+            self.can_interrupt,
+            self.expected_layout,
+            self.output_sample_rate,
+            self.output_frame_size,
+            self.input_sample_rate,
+            self.model,
+        )

backend/fastrtc/speech_to_text/__init__.py

@@ -0,0 +1,3 @@
+from .stt_ import MoonshineSTT, get_stt_model, stt_for_chunks
+
+__all__ = ["get_stt_model", "MoonshineSTT", "get_stt_model", "stt_for_chunks"]

backend/fastrtc/speech_to_text/stt_.py

@@ -0,0 +1,76 @@
+from functools import lru_cache
+from pathlib import Path
+from typing import Literal, Protocol
+
+import click
+import librosa
+import numpy as np
+from numpy.typing import NDArray
+
+from ..utils import AudioChunk, audio_to_float32
+
+curr_dir = Path(__file__).parent
+
+
+class STTModel(Protocol):
+    def stt(self, audio: tuple[int, NDArray[np.int16 | np.float32]]) -> str: ...
+
+
+class MoonshineSTT(STTModel):
+    def __init__(
+        self, model: Literal["moonshine/base", "moonshine/tiny"] = "moonshine/base"
+    ):
+        try:
+            from moonshine_onnx import MoonshineOnnxModel, load_tokenizer
+        except (ImportError, ModuleNotFoundError):
+            raise ImportError(
+                "Install fastrtc[stt] for speech-to-text and stopword detection support."
+            )
+
+        self.model = MoonshineOnnxModel(model_name=model)
+        self.tokenizer = load_tokenizer()
+
+    def stt(self, audio: tuple[int, NDArray[np.int16 | np.float32]]) -> str:
+        sr, audio_np = audio  # type: ignore
+        if audio_np.dtype == np.int16:
+            audio_np = audio_to_float32(audio)
+        if sr != 16000:
+            audio_np: NDArray[np.float32] = librosa.resample(
+                audio_np, orig_sr=sr, target_sr=16000
+            )
+        if audio_np.ndim == 1:
+            audio_np = audio_np.reshape(1, -1)
+        tokens = self.model.generate(audio_np)
+        return self.tokenizer.decode_batch(tokens)[0]
+
+
+@lru_cache
+def get_stt_model(
+    model: Literal["moonshine/base", "moonshine/tiny"] = "moonshine/base",
+) -> STTModel:
+    import os
+
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    m = MoonshineSTT(model)
+    from moonshine_onnx import load_audio
+
+    audio = load_audio(str(curr_dir / "test_file.wav"))
+    print(click.style("INFO", fg="green") + ":\t  Warming up STT model.")
+
+    m.stt((16000, audio))
+    print(click.style("INFO", fg="green") + ":\t  STT model warmed up.")
+    return m
+
+
+def stt_for_chunks(
+    stt_model: STTModel,
+    audio: tuple[int, NDArray[np.int16 | np.float32]],
+    chunks: list[AudioChunk],
+) -> str:
+    sr, audio_np = audio
+    return " ".join(
+        [
+            stt_model.stt((sr, audio_np[chunk["start"] : chunk["end"]]))
+            for chunk in chunks
+        ]
+    )

backend/fastrtc/stream.py

@@ -0,0 +1,721 @@
+import logging
+from pathlib import Path
+from typing import (
+    Any,
+    AsyncContextManager,
+    Callable,
+    Literal,
+    TypedDict,
+    cast,
+)
+
+import gradio as gr
+from fastapi import FastAPI, Request, WebSocket
+from fastapi.responses import HTMLResponse
+from gradio import Blocks
+from gradio.components.base import Component
+from pydantic import BaseModel
+from typing_extensions import NotRequired
+
+from .tracks import HandlerType, StreamHandlerImpl
+from .webrtc import WebRTC
+from .webrtc_connection_mixin import WebRTCConnectionMixin
+from .websocket import WebSocketHandler
+
+logger = logging.getLogger(__name__)
+
+curr_dir = Path(__file__).parent
+
+
+class Body(BaseModel):
+    sdp: str
+    type: str
+    webrtc_id: str
+
+
+class UIArgs(TypedDict):
+    title: NotRequired[str]
+    """Title of the demo"""
+    subtitle: NotRequired[str]
+    """Subtitle of the demo. Text will be centered and displayed below the title."""
+    icon: NotRequired[str]
+    """Icon to display on the button instead of the wave animation. The icon should be a path/url to a .svg/.png/.jpeg file."""
+    icon_button_color: NotRequired[str]
+    """Color of the icon button. Default is var(--color-accent) of the demo theme."""
+    pulse_color: NotRequired[str]
+    """Color of the pulse animation. Default is var(--color-accent) of the demo theme."""
+    icon_radius: NotRequired[int]
+    """Border radius of the icon button expressed as a percentage of the button size. Default is 50%."""
+
+
+class Stream(WebRTCConnectionMixin):
+    def __init__(
+        self,
+        handler: HandlerType,
+        *,
+        additional_outputs_handler: Callable | None = None,
+        mode: Literal["send-receive", "receive", "send"] = "send-receive",
+        modality: Literal["video", "audio", "audio-video"] = "video",
+        concurrency_limit: int | None | Literal["default"] = "default",
+        time_limit: float | None = None,
+        rtp_params: dict[str, Any] | None = None,
+        rtc_configuration: dict[str, Any] | None = None,
+        additional_inputs: list[Component] | None = None,
+        additional_outputs: list[Component] | None = None,
+        ui_args: UIArgs | None = None,
+    ):
+        WebRTCConnectionMixin.__init__(self)
+        self.mode = mode
+        self.modality = modality
+        self.rtp_params = rtp_params
+        self.event_handler = handler
+        self.concurrency_limit = cast(
+            (int | float),
+            1 if concurrency_limit in ["default", None] else concurrency_limit,
+        )
+        self.time_limit = time_limit
+        self.additional_output_components = additional_outputs
+        self.additional_input_components = additional_inputs
+        self.additional_outputs_handler = additional_outputs_handler
+        self.rtc_configuration = rtc_configuration
+        self._ui = self._generate_default_ui(ui_args)
+        self._ui.launch = self._wrap_gradio_launch(self._ui.launch)
+
+    def mount(self, app: FastAPI):
+        app.router.post("/webrtc/offer")(self.offer)
+        app.router.websocket("/telephone/handler")(self.telephone_handler)
+        app.router.post("/telephone/incoming")(self.handle_incoming_call)
+        app.router.websocket("/websocket/offer")(self.websocket_offer)
+        lifespan = self._inject_startup_message(app.router.lifespan_context)
+        app.router.lifespan_context = lifespan
+
+    @staticmethod
+    def print_error(env: Literal["colab", "spaces"]):
+        import click
+
+        print(
+            click.style("ERROR", fg="red")
+            + f":\t  Running in {env} is not possible without providing a valid rtc_configuration. "
+            + "See "
+            + click.style("https://fastrtc.org/deployment/", fg="cyan")
+            + " for more information."
+        )
+        raise RuntimeError(
+            f"Running in {env} is not possible without providing a valid rtc_configuration. "
+            + "See https://fastrtc.org/deployment/ for more information."
+        )
+
+    def _check_colab_or_spaces(self):
+        from gradio.utils import colab_check, get_space
+
+        if colab_check() and not self.rtc_configuration:
+            self.print_error("colab")
+        if get_space() and not self.rtc_configuration:
+            self.print_error("spaces")
+
+    def _wrap_gradio_launch(self, callable):
+        import contextlib
+
+        def wrapper(*args, **kwargs):
+            lifespan = kwargs.get("app_kwargs", {}).get("lifespan", None)
+
+            @contextlib.asynccontextmanager
+            async def new_lifespan(app: FastAPI):
+                if lifespan is None:
+                    self._check_colab_or_spaces()
+                    yield
+                else:
+                    async with lifespan(app):
+                        self._check_colab_or_spaces()
+                        yield
+
+            if "app_kwargs" not in kwargs:
+                kwargs["app_kwargs"] = {}
+            kwargs["app_kwargs"]["lifespan"] = new_lifespan
+            return callable(*args, **kwargs)
+
+        return wrapper
+
+    def _inject_startup_message(
+        self, lifespan: Callable[[FastAPI], AsyncContextManager] | None = None
+    ):
+        import contextlib
+
+        import click
+
+        def print_startup_message():
+            self._check_colab_or_spaces()
+            print(
+                click.style("INFO", fg="green")
+                + ":\t  Visit "
+                + click.style("https://fastrtc.org/userguide/api/", fg="cyan")
+                + " for WebRTC or Websocket API docs."
+            )
+
+        @contextlib.asynccontextmanager
+        async def new_lifespan(app: FastAPI):
+            if lifespan is None:
+                print_startup_message()
+                yield
+            else:
+                async with lifespan(app):
+                    print_startup_message()
+                    yield
+
+        return new_lifespan
+
+    def _generate_default_ui(
+        self,
+        ui_args: UIArgs | None = None,
+    ):
+        ui_args = ui_args or {}
+        same_components = []
+        additional_input_components = self.additional_input_components or []
+        additional_output_components = self.additional_output_components or []
+        if additional_output_components and not self.additional_outputs_handler:
+            raise ValueError(
+                "additional_outputs_handler must be provided if there are additional output components."
+            )
+        if additional_input_components and additional_output_components:
+            same_components = [
+                component
+                for component in additional_input_components
+                if component in additional_output_components
+            ]
+            for component in additional_output_components:
+                if component in same_components:
+                    same_components.append(component)
+        if self.modality == "video" and self.mode == "receive":
+            with gr.Blocks() as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Video Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Row():
+                    with gr.Column():
+                        if additional_input_components:
+                            for component in additional_input_components:
+                                component.render()
+                        button = gr.Button("Start Stream", variant="primary")
+                    with gr.Column():
+                        output_video = WebRTC(
+                            label="Video Stream",
+                            rtc_configuration=self.rtc_configuration,
+                            mode="receive",
+                            modality="video",
+                        )
+                        for component in additional_output_components:
+                            if component not in same_components:
+                                component.render()
+                output_video.stream(
+                    fn=self.event_handler,
+                    inputs=self.additional_input_components,
+                    outputs=[output_video],
+                    trigger=button.click,
+                    time_limit=self.time_limit,
+                    concurrency_limit=self.concurrency_limit,  # type: ignore
+                )
+                if additional_output_components:
+                    assert self.additional_outputs_handler
+                    output_video.on_additional_outputs(
+                        self.additional_outputs_handler,
+                        inputs=additional_output_components,
+                        outputs=additional_output_components,
+                    )
+        elif self.modality == "video" and self.mode == "send":
+            with gr.Blocks() as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Video Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Row():
+                    if additional_input_components:
+                        with gr.Column():
+                            for component in additional_input_components:
+                                component.render()
+                    with gr.Column():
+                        output_video = WebRTC(
+                            label="Video Stream",
+                            rtc_configuration=self.rtc_configuration,
+                            mode="send",
+                            modality="video",
+                        )
+                        for component in additional_output_components:
+                            if component not in same_components:
+                                component.render()
+                output_video.stream(
+                    fn=self.event_handler,
+                    inputs=[output_video] + additional_input_components,
+                    outputs=[output_video],
+                    time_limit=self.time_limit,
+                    concurrency_limit=self.concurrency_limit,  # type: ignore
+                )
+                if additional_output_components:
+                    assert self.additional_outputs_handler
+                    output_video.on_additional_outputs(
+                        self.additional_outputs_handler,
+                        inputs=additional_output_components,
+                        outputs=additional_output_components,
+                    )
+        elif self.modality == "video" and self.mode == "send-receive":
+            css = """.my-group {max-width: 600px !important; max-height: 600 !important;}
+                      .my-column {display: flex !important; justify-content: center !important; align-items: center !important};"""
+
+            with gr.Blocks(css=css) as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Video Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Column(elem_classes=["my-column"]):
+                    with gr.Group(elem_classes=["my-group"]):
+                        image = WebRTC(
+                            label="Stream",
+                            rtc_configuration=self.rtc_configuration,
+                            mode="send-receive",
+                            modality="video",
+                        )
+                        for component in additional_input_components:
+                            component.render()
+                if additional_output_components:
+                    with gr.Column():
+                        for component in additional_output_components:
+                            if component not in same_components:
+                                component.render()
+
+                image.stream(
+                    fn=self.event_handler,
+                    inputs=[image] + additional_input_components,
+                    outputs=[image],
+                    time_limit=self.time_limit,
+                    concurrency_limit=self.concurrency_limit,  # type: ignore
+                )
+                if additional_output_components:
+                    assert self.additional_outputs_handler
+                    image.on_additional_outputs(
+                        self.additional_outputs_handler,
+                        inputs=additional_output_components,
+                        outputs=additional_output_components,
+                    )
+        elif self.modality == "audio" and self.mode == "receive":
+            with gr.Blocks() as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Audio Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Row():
+                    with gr.Column():
+                        for component in additional_input_components:
+                            component.render()
+                        button = gr.Button("Start Stream", variant="primary")
+                    if additional_output_components:
+                        with gr.Column():
+                            output_video = WebRTC(
+                                label="Audio Stream",
+                                rtc_configuration=self.rtc_configuration,
+                                mode="receive",
+                                modality="audio",
+                                icon=ui_args.get("icon"),
+                                icon_button_color=ui_args.get("icon_button_color"),
+                                pulse_color=ui_args.get("pulse_color"),
+                                icon_radius=ui_args.get("icon_radius"),
+                            )
+                            for component in additional_output_components:
+                                if component not in same_components:
+                                    component.render()
+                output_video.stream(
+                    fn=self.event_handler,
+                    inputs=self.additional_input_components,
+                    outputs=[output_video],
+                    trigger=button.click,
+                    time_limit=self.time_limit,
+                    concurrency_limit=self.concurrency_limit,  # type: ignore
+                )
+                if additional_output_components:
+                    assert self.additional_outputs_handler
+                    output_video.on_additional_outputs(
+                        self.additional_outputs_handler,
+                        inputs=additional_output_components,
+                        outputs=additional_output_components,
+                    )
+        elif self.modality == "audio" and self.mode == "send":
+            with gr.Blocks() as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Audio Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Row():
+                    with gr.Column():
+                        with gr.Group():
+                            image = WebRTC(
+                                label="Stream",
+                                rtc_configuration=self.rtc_configuration,
+                                mode="send",
+                                modality="audio",
+                                icon=ui_args.get("icon"),
+                                icon_button_color=ui_args.get("icon_button_color"),
+                                pulse_color=ui_args.get("pulse_color"),
+                                icon_radius=ui_args.get("icon_radius"),
+                            )
+                            for component in additional_input_components:
+                                if component not in same_components:
+                                    component.render()
+                    if additional_output_components:
+                        with gr.Column():
+                            for component in additional_output_components:
+                                component.render()
+                image.stream(
+                    fn=self.event_handler,
+                    inputs=[image] + additional_input_components,
+                    outputs=[image],
+                    time_limit=self.time_limit,
+                    concurrency_limit=self.concurrency_limit,  # type: ignore
+                )
+                if additional_output_components:
+                    assert self.additional_outputs_handler
+                    image.on_additional_outputs(
+                        self.additional_outputs_handler,
+                        inputs=additional_output_components,
+                        outputs=additional_output_components,
+                    )
+        elif self.modality == "audio" and self.mode == "send-receive":
+            with gr.Blocks() as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Audio Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                    {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Row():
+                    with gr.Column():
+                        with gr.Group():
+                            image = WebRTC(
+                                label="Stream",
+                                rtc_configuration=self.rtc_configuration,
+                                mode="send-receive",
+                                modality="audio",
+                                icon=ui_args.get("icon"),
+                                icon_button_color=ui_args.get("icon_button_color"),
+                                pulse_color=ui_args.get("pulse_color"),
+                                icon_radius=ui_args.get("icon_radius"),
+                            )
+                            for component in additional_input_components:
+                                if component not in same_components:
+                                    component.render()
+                    if additional_output_components:
+                        with gr.Column():
+                            for component in additional_output_components:
+                                component.render()
+
+                    image.stream(
+                        fn=self.event_handler,
+                        inputs=[image] + additional_input_components,
+                        outputs=[image],
+                        time_limit=self.time_limit,
+                        concurrency_limit=self.concurrency_limit,  # type: ignore
+                    )
+                    if additional_output_components:
+                        assert self.additional_outputs_handler
+                        image.on_additional_outputs(
+                            self.additional_outputs_handler,
+                            inputs=additional_output_components,
+                            outputs=additional_output_components,
+                        )
+        elif self.modality == "audio-video" and self.mode == "send-receive":
+            with gr.Blocks() as demo:
+                gr.HTML(
+                    f"""
+                <h1 style='text-align: center'>
+                {ui_args.get("title", "Audio Video Streaming (Powered by FastRTC ⚡️)")}
+                </h1>
+                """
+                )
+                if ui_args.get("subtitle"):
+                    gr.Markdown(
+                        f"""
+                <div style='text-align: center'>
+                {ui_args.get("subtitle")}
+                </div>
+                """
+                    )
+                with gr.Row():
+                    with gr.Column():
+                        with gr.Group():
+                            image = WebRTC(
+                                label="Stream",
+                                rtc_configuration=self.rtc_configuration,
+                                mode="send-receive",
+                                modality="audio-video",
+                                icon=ui_args.get("icon"),
+                                icon_button_color=ui_args.get("icon_button_color"),
+                                pulse_color=ui_args.get("pulse_color"),
+                                icon_radius=ui_args.get("icon_radius"),
+                            )
+                            for component in additional_input_components:
+                                if component not in same_components:
+                                    component.render()
+                    if additional_output_components:
+                        with gr.Column():
+                            for component in additional_output_components:
+                                component.render()
+
+                    image.stream(
+                        fn=self.event_handler,
+                        inputs=[image] + additional_input_components,
+                        outputs=[image],
+                        time_limit=self.time_limit,
+                        concurrency_limit=self.concurrency_limit,  # type: ignore
+                    )
+                    if additional_output_components:
+                        assert self.additional_outputs_handler
+                        image.on_additional_outputs(
+                            self.additional_outputs_handler,
+                            inputs=additional_output_components,
+                            outputs=additional_output_components,
+                        )
+        else:
+            raise ValueError(f"Invalid modality: {self.modality} and mode: {self.mode}")
+        return demo
+
+    @property
+    def ui(self) -> Blocks:
+        return self._ui
+
+    @ui.setter
+    def ui(self, blocks: Blocks):
+        self._ui = blocks
+
+    async def offer(self, body: Body):
+        return await self.handle_offer(
+            body.model_dump(), set_outputs=self.set_additional_outputs(body.webrtc_id)
+        )
+
+    async def handle_incoming_call(self, request: Request):
+        from twilio.twiml.voice_response import Connect, VoiceResponse
+
+        response = VoiceResponse()
+        response.say("Connecting to the AI assistant.")
+        connect = Connect()
+        connect.stream(url=f"wss://{request.url.hostname}/telephone/handler")
+        response.append(connect)
+        response.say("The call has been disconnected.")
+        return HTMLResponse(content=str(response), media_type="application/xml")
+
+    async def telephone_handler(self, websocket: WebSocket):
+        handler = cast(StreamHandlerImpl, self.event_handler.copy())  # type: ignore
+        handler.phone_mode = True
+
+        async def set_handler(s: str, a: WebSocketHandler):
+            if len(self.connections) >= self.concurrency_limit:  # type: ignore
+                await cast(WebSocket, a.websocket).send_json(
+                    {
+                        "status": "failed",
+                        "meta": {
+                            "error": "concurrency_limit_reached",
+                            "limit": self.concurrency_limit,
+                        },
+                    }
+                )
+                await websocket.close()
+                return
+
+        ws = WebSocketHandler(
+            handler, set_handler, lambda s: None, lambda s: lambda a: None
+        )
+        await ws.handle_websocket(websocket)
+
+    async def websocket_offer(self, websocket: WebSocket):
+        handler = cast(StreamHandlerImpl, self.event_handler.copy())  # type: ignore
+        handler.phone_mode = False
+
+        async def set_handler(s: str, a: WebSocketHandler):
+            if len(self.connections) >= self.concurrency_limit:  # type: ignore
+                await cast(WebSocket, a.websocket).send_json(
+                    {
+                        "status": "failed",
+                        "meta": {
+                            "error": "concurrency_limit_reached",
+                            "limit": self.concurrency_limit,
+                        },
+                    }
+                )
+                await websocket.close()
+                return
+
+            self.connections[s] = [a]  # type: ignore
+
+        def clean_up(s):
+            self.clean_up(s)
+
+        ws = WebSocketHandler(
+            handler, set_handler, clean_up, lambda s: self.set_additional_outputs(s)
+        )
+        await ws.handle_websocket(websocket)
+
+    def fastphone(
+        self,
+        token: str | None = None,
+        host: str = "127.0.0.1",
+        port: int = 8000,
+        **kwargs,
+    ):
+        import atexit
+        import secrets
+        import threading
+        import time
+        import urllib.parse
+
+        import click
+        import httpx
+        import uvicorn
+        from gradio.networking import setup_tunnel
+        from gradio.tunneling import CURRENT_TUNNELS
+        from huggingface_hub import get_token
+
+        app = FastAPI()
+
+        self.mount(app)
+
+        t = threading.Thread(
+            target=uvicorn.run,
+            args=(app,),
+            kwargs={"host": host, "port": port, **kwargs},
+        )
+        t.start()
+
+        url = setup_tunnel(
+            host, port, share_token=secrets.token_urlsafe(32), share_server_address=None
+        )
+        host = urllib.parse.urlparse(url).netloc
+
+        URL = "https://api.fastrtc.org"
+        try:
+            r = httpx.post(
+                URL + "/register",
+                json={"url": host},
+                headers={"Authorization": token or get_token() or ""},
+            )
+        except Exception:
+            URL = "https://fastrtc-fastphone.hf.space"
+            r = httpx.post(
+                URL + "/register",
+                json={"url": host},
+                headers={"Authorization": token or get_token() or ""},
+            )
+        r.raise_for_status()
+        data = r.json()
+        code = f"{data['code']}"
+        phone_number = data["phone"]
+        reset_date = data["reset_date"]
+        print(
+            click.style("INFO", fg="green")
+            + ":\t  Your FastPhone is now live! Call "
+            + click.style(phone_number, fg="cyan")
+            + " and use code "
+            + click.style(code, fg="cyan")
+            + " to connect to your stream."
+        )
+        minutes = str(int(data["time_remaining"] // 60)).zfill(2)
+        seconds = str(int(data["time_remaining"] % 60)).zfill(2)
+        print(
+            click.style("INFO", fg="green")
+            + ":\t  You have "
+            + click.style(f"{minutes}:{seconds}", fg="cyan")
+            + " minutes remaining in your quota (Resetting on "
+            + click.style(f"{reset_date}", fg="cyan")
+            + ")"
+        )
+        print(
+            click.style("INFO", fg="green")
+            + ":\t  Visit "
+            + click.style(
+                "https://fastrtc.org/userguide/audio/#telephone-integration",
+                fg="cyan",
+            )
+            + " for information on making your handler compatible with phone usage."
+        )
+
+        def unregister():
+            httpx.post(
+                URL + "/unregister",
+                json={"url": host, "code": code},
+                headers={"Authorization": token or get_token() or ""},
+            )
+
+        atexit.register(unregister)
+
+        try:
+            while True:
+                time.sleep(0.1)
+        except (KeyboardInterrupt, OSError):
+            print(
+                click.style("INFO", fg="green")
+                + ":\t  Keyboard interruption in main thread... closing server."
+            )
+            unregister()
+            t.join(timeout=5)
+            for tunnel in CURRENT_TUNNELS:
+                tunnel.kill()

backend/fastrtc/text_to_speech/__init__.py

@@ -0,0 +1,3 @@
+from .tts import KokoroTTSOptions, get_tts_model
+
+__all__ = ["get_tts_model", "KokoroTTSOptions"]

backend/fastrtc/text_to_speech/test_tts.py

@@ -0,0 +1,13 @@
+from fastrtc.text_to_speech.tts import get_tts_model
+
+
+def test_tts_long_prompt():
+    model = get_tts_model()
+    prompt = "It may be that this communication will be considered as a madman's freak but at any rate it must be admitted that in its clearness and frankness it left nothing to be desired The serious part of it was that the Federal Government had undertaken to treat a sale by auction as a valid concession of these undiscovered territories Opinions on the matter were many Some readers saw in it only one of those prodigious outbursts of American humbug which would exceed the limits of puffism if the depths of human credulity were not unfathomable"
+
+    for i, chunk in enumerate(model.stream_tts_sync(prompt)):
+        print(f"Chunk {i}: {chunk[1].shape}")
+
+
+if __name__ == "__main__":
+    test_tts_long_prompt()

backend/fastrtc/text_to_speech/tts.py

@@ -0,0 +1,135 @@
+import asyncio
+import re
+from dataclasses import dataclass
+from functools import lru_cache
+from typing import AsyncGenerator, Generator, Literal, Protocol
+
+import numpy as np
+from huggingface_hub import hf_hub_download
+from numpy.typing import NDArray
+
+
+class TTSOptions:
+    pass
+
+
+class TTSModel(Protocol):
+    def tts(self, text: str) -> tuple[int, NDArray[np.float32]]: ...
+
+    async def stream_tts(
+        self, text: str, options: TTSOptions | None = None
+    ) -> AsyncGenerator[tuple[int, NDArray[np.float32]], None]: ...
+
+    def stream_tts_sync(
+        self, text: str, options: TTSOptions | None = None
+    ) -> Generator[tuple[int, NDArray[np.float32]], None, None]: ...
+
+
+@dataclass
+class KokoroTTSOptions(TTSOptions):
+    voice: str = "af_heart"
+    speed: float = 1.0
+    lang: str = "en-us"
+
+
+@lru_cache
+def get_tts_model(model: Literal["kokoro"] = "kokoro") -> TTSModel:
+    m = KokoroTTSModel()
+    m.tts("Hello, world!")
+    return m
+
+
+class KokoroFixedBatchSize:
+    # Source: https://github.com/thewh1teagle/kokoro-onnx/issues/115#issuecomment-2676625392
+    def _split_phonemes(self, phonemes: str) -> list[str]:
+        MAX_PHONEME_LENGTH = 510
+        max_length = MAX_PHONEME_LENGTH - 1
+        batched_phonemes = []
+        while len(phonemes) > max_length:
+            # Find best split point within limit
+            split_idx = max_length
+
+            # Try to find the last period before max_length
+            period_idx = phonemes.rfind(".", 0, max_length)
+            if period_idx != -1:
+                split_idx = period_idx + 1  # Include period
+
+            else:
+                # Try other punctuation
+                match = re.search(
+                    r"[!?;,]", phonemes[:max_length][::-1]
+                )  # Search backwards
+                if match:
+                    split_idx = max_length - match.start()
+
+                else:
+                    # Try last space
+                    space_idx = phonemes.rfind(" ", 0, max_length)
+                    if space_idx != -1:
+                        split_idx = space_idx
+
+            # If no good split point is found, force split at max_length
+            chunk = phonemes[:split_idx].strip()
+            batched_phonemes.append(chunk)
+
+            # Move to the next part
+            phonemes = phonemes[split_idx:].strip()
+
+        # Add remaining phonemes
+        if phonemes:
+            batched_phonemes.append(phonemes)
+        return batched_phonemes
+
+
+class KokoroTTSModel(TTSModel):
+    def __init__(self):
+        from kokoro_onnx import Kokoro
+
+        self.model = Kokoro(
+            model_path=hf_hub_download("fastrtc/kokoro-onnx", "kokoro-v1.0.onnx"),
+            voices_path=hf_hub_download("fastrtc/kokoro-onnx", "voices-v1.0.bin"),
+        )
+
+        self.model._split_phonemes = KokoroFixedBatchSize()._split_phonemes
+
+    def tts(
+        self, text: str, options: KokoroTTSOptions | None = None
+    ) -> tuple[int, NDArray[np.float32]]:
+        options = options or KokoroTTSOptions()
+        a, b = self.model.create(
+            text, voice=options.voice, speed=options.speed, lang=options.lang
+        )
+        return b, a
+
+    async def stream_tts(
+        self, text: str, options: KokoroTTSOptions | None = None
+    ) -> AsyncGenerator[tuple[int, NDArray[np.float32]], None]:
+        options = options or KokoroTTSOptions()
+
+        sentences = re.split(r"(?<=[.!?])\s+", text.strip())
+
+        for s_idx, sentence in enumerate(sentences):
+            if not sentence.strip():
+                continue
+
+            chunk_idx = 0
+            async for chunk in self.model.create_stream(
+                sentence, voice=options.voice, speed=options.speed, lang=options.lang
+            ):
+                if s_idx != 0 and chunk_idx == 0:
+                    yield chunk[1], np.zeros(chunk[1] // 7, dtype=np.float32)
+                chunk_idx += 1
+                yield chunk[1], chunk[0]
+
+    def stream_tts_sync(
+        self, text: str, options: KokoroTTSOptions | None = None
+    ) -> Generator[tuple[int, NDArray[np.float32]], None, None]:
+        loop = asyncio.new_event_loop()
+
+        # Use the new loop to run the async generator
+        iterator = self.stream_tts(text, options).__aiter__()
+        while True:
+            try:
+                yield loop.run_until_complete(iterator.__anext__())
+            except StopAsyncIteration:
+                break

backend/fastrtc/tracks.py

@@ -0,0 +1,731 @@
+"""WebRTC tracks."""
+
+from __future__ import annotations
+
+import asyncio
+import functools
+import inspect
+import logging
+import threading
+import time
+import traceback
+from abc import ABC, abstractmethod
+from collections.abc import Callable
+from typing import (
+    Any,
+    Generator,
+    Literal,
+    TypeAlias,
+    Union,
+    cast,
+)
+
+import anyio.to_thread
+import av
+import numpy as np
+from aiortc import (
+    AudioStreamTrack,
+    MediaStreamTrack,
+    VideoStreamTrack,
+)
+from aiortc.contrib.media import AudioFrame, VideoFrame  # type: ignore
+from aiortc.mediastreams import MediaStreamError
+from numpy import typing as npt
+
+from fastrtc.utils import (
+    AdditionalOutputs,
+    DataChannel,
+    WebRTCError,
+    create_message,
+    current_channel,
+    player_worker_decode,
+    split_output,
+)
+
+logger = logging.getLogger(__name__)
+
+VideoNDArray: TypeAlias = Union[
+    np.ndarray[Any, np.dtype[np.uint8]],
+    np.ndarray[Any, np.dtype[np.uint16]],
+    np.ndarray[Any, np.dtype[np.float32]],
+]
+
+VideoEmitType = (
+    VideoNDArray | tuple[VideoNDArray, AdditionalOutputs] | AdditionalOutputs
+)
+VideoEventHandler = Callable[[npt.ArrayLike], VideoEmitType]
+
+
+class VideoCallback(VideoStreamTrack):
+    """
+    This works for streaming input and output
+    """
+
+    kind = "video"
+
+    def __init__(
+        self,
+        track: MediaStreamTrack,
+        event_handler: VideoEventHandler,
+        channel: DataChannel | None = None,
+        set_additional_outputs: Callable | None = None,
+        mode: Literal["send-receive", "send"] = "send-receive",
+    ) -> None:
+        super().__init__()  # don't forget this!
+        self.track = track
+        self.event_handler = event_handler
+        self.latest_args: str | list[Any] = "not_set"
+        self.channel = channel
+        self.set_additional_outputs = set_additional_outputs
+        self.thread_quit = asyncio.Event()
+        self.mode = mode
+        self.channel_set = asyncio.Event()
+        self.has_started = False
+
+    def set_channel(self, channel: DataChannel):
+        self.channel = channel
+        current_channel.set(channel)
+        self.channel_set.set()
+
+    def set_args(self, args: list[Any]):
+        self.latest_args = ["__webrtc_value__"] + list(args)
+
+    def add_frame_to_payload(
+        self, args: list[Any], frame: np.ndarray | None
+    ) -> list[Any]:
+        new_args = []
+        for val in args:
+            if isinstance(val, str) and val == "__webrtc_value__":
+                new_args.append(frame)
+            else:
+                new_args.append(val)
+        return new_args
+
+    def array_to_frame(self, array: np.ndarray) -> VideoFrame:
+        return VideoFrame.from_ndarray(array, format="bgr24")
+
+    async def process_frames(self):
+        while not self.thread_quit.is_set():
+            try:
+                await self.recv()
+            except TimeoutError:
+                continue
+
+    async def start(
+        self,
+    ):
+        asyncio.create_task(self.process_frames())
+
+    def stop(self):
+        super().stop()
+        logger.debug("video callback stop")
+        self.thread_quit.set()
+
+    async def wait_for_channel(self):
+        if not self.channel_set.is_set():
+            await self.channel_set.wait()
+        if current_channel.get() != self.channel:
+            current_channel.set(self.channel)
+
+    async def recv(self):  # type: ignore
+        try:
+            try:
+                frame = cast(VideoFrame, await self.track.recv())
+            except MediaStreamError:
+                self.stop()
+                return
+
+            await self.wait_for_channel()
+            frame_array = frame.to_ndarray(format="bgr24")
+            if self.latest_args == "not_set":
+                return frame
+
+            args = self.add_frame_to_payload(cast(list, self.latest_args), frame_array)
+
+            array, outputs = split_output(self.event_handler(*args))
+            if (
+                isinstance(outputs, AdditionalOutputs)
+                and self.set_additional_outputs
+                and self.channel
+            ):
+                self.set_additional_outputs(outputs)
+                self.channel.send(create_message("fetch_output", []))
+            if array is None and self.mode == "send":
+                return
+
+            new_frame = self.array_to_frame(array)
+            if frame:
+                new_frame.pts = frame.pts
+                new_frame.time_base = frame.time_base
+            else:
+                pts, time_base = await self.next_timestamp()
+                new_frame.pts = pts
+                new_frame.time_base = time_base
+
+            return new_frame
+        except Exception as e:
+            logger.debug("exception %s", e)
+            exec = traceback.format_exc()
+            logger.debug("traceback %s", exec)
+            if isinstance(e, WebRTCError):
+                raise e
+            else:
+                raise WebRTCError(str(e)) from e
+
+
+class StreamHandlerBase(ABC):
+    def __init__(
+        self,
+        expected_layout: Literal["mono", "stereo"] = "mono",
+        output_sample_rate: int = 24000,
+        output_frame_size: int = 960,
+        input_sample_rate: int = 48000,
+    ) -> None:
+        self.expected_layout = expected_layout
+        self.output_sample_rate = output_sample_rate
+        self.output_frame_size = output_frame_size
+        self.input_sample_rate = input_sample_rate
+        self.latest_args: list[Any] = []
+        self._resampler = None
+        self._channel: DataChannel | None = None
+        self._loop: asyncio.AbstractEventLoop
+        self.args_set = asyncio.Event()
+        self.channel_set = asyncio.Event()
+        self._phone_mode = False
+        self._clear_queue: Callable | None = None
+
+    @property
+    def clear_queue(self) -> Callable:
+        return cast(Callable, self._clear_queue)
+
+    @property
+    def loop(self) -> asyncio.AbstractEventLoop:
+        return cast(asyncio.AbstractEventLoop, self._loop)
+
+    @property
+    def channel(self) -> DataChannel | None:
+        return self._channel
+
+    @property
+    def phone_mode(self) -> bool:
+        return self._phone_mode
+
+    @phone_mode.setter
+    def phone_mode(self, value: bool):
+        self._phone_mode = value
+
+    def set_channel(self, channel: DataChannel):
+        self._channel = channel
+        self.channel_set.set()
+
+    async def fetch_args(
+        self,
+    ):
+        if self.channel:
+            self.channel.send(create_message("send_input", []))
+            logger.debug("Sent send_input")
+
+    async def wait_for_args(self):
+        if not self.phone_mode:
+            await self.fetch_args()
+            await self.args_set.wait()
+        else:
+            self.args_set.set()
+
+    def wait_for_args_sync(self):
+        try:
+            asyncio.run_coroutine_threadsafe(self.wait_for_args(), self.loop).result()
+        except Exception:
+            import traceback
+
+            traceback.print_exc()
+
+    async def send_message(self, msg: str):
+        if self.channel:
+            self.channel.send(msg)
+            logger.debug("Sent msg %s", msg)
+
+    def send_message_sync(self, msg: str):
+        try:
+            asyncio.run_coroutine_threadsafe(self.send_message(msg), self.loop).result()
+            logger.debug("Sent msg %s", msg)
+        except Exception as e:
+            logger.debug("Exception sending msg %s", e)
+
+    def set_args(self, args: list[Any]):
+        logger.debug("setting args in audio callback %s", args)
+        self.latest_args = ["__webrtc_value__"] + list(args)
+        self.args_set.set()
+
+    def reset(self):
+        self.args_set.clear()
+
+    def shutdown(self):
+        pass
+
+    def resample(self, frame: AudioFrame) -> Generator[AudioFrame, None, None]:
+        if self._resampler is None:
+            self._resampler = av.AudioResampler(  # type: ignore
+                format="s16",
+                layout=self.expected_layout,
+                rate=self.input_sample_rate,
+                frame_size=frame.samples,
+            )
+        yield from self._resampler.resample(frame)
+
+
+EmitType: TypeAlias = (
+    tuple[int, npt.NDArray[np.int16 | np.float32]]
+    | tuple[int, npt.NDArray[np.int16 | np.float32], Literal["mono", "stereo"]]
+    | AdditionalOutputs
+    | tuple[tuple[int, npt.NDArray[np.int16 | np.float32]], AdditionalOutputs]
+    | None
+)
+AudioEmitType = EmitType
+
+
+class StreamHandler(StreamHandlerBase):
+    @abstractmethod
+    def receive(self, frame: tuple[int, npt.NDArray[np.int16]]) -> None:
+        pass
+
+    @abstractmethod
+    def emit(self) -> EmitType:
+        pass
+
+    @abstractmethod
+    def copy(self) -> StreamHandler:
+        pass
+
+    def start_up(self):
+        pass
+
+
+class AsyncStreamHandler(StreamHandlerBase):
+    @abstractmethod
+    async def receive(self, frame: tuple[int, npt.NDArray[np.int16]]) -> None:
+        pass
+
+    @abstractmethod
+    async def emit(self) -> EmitType:
+        pass
+
+    @abstractmethod
+    def copy(self) -> AsyncStreamHandler:
+        pass
+
+    async def start_up(self):
+        pass
+
+
+StreamHandlerImpl = StreamHandler | AsyncStreamHandler
+
+
+class AudioVideoStreamHandler(StreamHandler):
+    @abstractmethod
+    def video_receive(self, frame: VideoFrame) -> None:
+        pass
+
+    @abstractmethod
+    def video_emit(self) -> VideoEmitType:
+        pass
+
+    @abstractmethod
+    def copy(self) -> AudioVideoStreamHandler:
+        pass
+
+
+class AsyncAudioVideoStreamHandler(AsyncStreamHandler):
+    @abstractmethod
+    async def video_receive(self, frame: npt.NDArray[np.float32]) -> None:
+        pass
+
+    @abstractmethod
+    async def video_emit(self) -> VideoEmitType:
+        pass
+
+    @abstractmethod
+    def copy(self) -> AsyncAudioVideoStreamHandler:
+        pass
+
+
+VideoStreamHandlerImpl = AudioVideoStreamHandler | AsyncAudioVideoStreamHandler
+AudioVideoStreamHandlerImpl = AudioVideoStreamHandler | AsyncAudioVideoStreamHandler
+AsyncHandler = AsyncStreamHandler | AsyncAudioVideoStreamHandler
+
+HandlerType = StreamHandlerImpl | VideoStreamHandlerImpl | VideoEventHandler | Callable
+
+
+class VideoStreamHandler(VideoCallback):
+    async def process_frames(self):
+        while not self.thread_quit.is_set():
+            try:
+                await self.channel_set.wait()
+                frame = cast(VideoFrame, await self.track.recv())
+                frame_array = frame.to_ndarray(format="bgr24")
+                handler = cast(VideoStreamHandlerImpl, self.event_handler)
+                if inspect.iscoroutinefunction(handler.video_receive):
+                    await handler.video_receive(frame_array)
+                else:
+                    handler.video_receive(frame_array)  # type: ignore
+            except MediaStreamError:
+                self.stop()
+
+    async def start(self):
+        if not self.has_started:
+            asyncio.create_task(self.process_frames())
+            self.has_started = True
+
+    async def recv(self):  # type: ignore
+        await self.start()
+        try:
+            handler = cast(VideoStreamHandlerImpl, self.event_handler)
+            if inspect.iscoroutinefunction(handler.video_emit):
+                outputs = await handler.video_emit()
+            else:
+                outputs = handler.video_emit()
+
+            array, outputs = split_output(outputs)
+            if (
+                isinstance(outputs, AdditionalOutputs)
+                and self.set_additional_outputs
+                and self.channel
+            ):
+                self.set_additional_outputs(outputs)
+                self.channel.send(create_message("fetch_output", []))
+            if array is None and self.mode == "send":
+                return
+
+            new_frame = self.array_to_frame(array)
+
+            # Will probably have to give developer ability to set pts and time_base
+            pts, time_base = await self.next_timestamp()
+            new_frame.pts = pts
+            new_frame.time_base = time_base
+
+            return new_frame
+        except Exception as e:
+            logger.debug("exception %s", e)
+            exec = traceback.format_exc()
+            logger.debug("traceback %s", exec)
+
+
+class AudioCallback(AudioStreamTrack):
+    kind = "audio"
+
+    def __init__(
+        self,
+        track: MediaStreamTrack,
+        event_handler: StreamHandlerBase,
+        channel: DataChannel | None = None,
+        set_additional_outputs: Callable | None = None,
+    ) -> None:
+        super().__init__()
+        self.track = track
+        self.event_handler = cast(StreamHandlerImpl, event_handler)
+        self.event_handler._clear_queue = self.clear_queue
+        self.current_timestamp = 0
+        self.latest_args: str | list[Any] = "not_set"
+        self.queue = asyncio.Queue()
+        self.thread_quit = asyncio.Event()
+        self._start: float | None = None
+        self.has_started = False
+        self.last_timestamp = 0
+        self.channel = channel
+        self.set_additional_outputs = set_additional_outputs
+
+    def clear_queue(self):
+        logger.debug("clearing queue")
+        logger.debug("queue size: %d", self.queue.qsize())
+        i = 0
+        while not self.queue.empty():
+            self.queue.get_nowait()
+            i += 1
+        logger.debug("popped %d items from queue", i)
+        self._start = None
+
+    async def wait_for_channel(self):
+        if not self.event_handler.channel_set.is_set():
+            await self.event_handler.channel_set.wait()
+        if current_channel.get() != self.event_handler.channel:
+            current_channel.set(self.event_handler.channel)
+
+    def set_channel(self, channel: DataChannel):
+        self.channel = channel
+        self.event_handler.set_channel(channel)
+
+    def set_args(self, args: list[Any]):
+        self.event_handler.set_args(args)
+
+    def event_handler_receive(self, frame: tuple[int, np.ndarray]) -> None:
+        current_channel.set(self.event_handler.channel)
+        return cast(Callable, self.event_handler.receive)(frame)
+
+    def event_handler_emit(self) -> EmitType:
+        current_channel.set(self.event_handler.channel)
+        return cast(Callable, self.event_handler.emit)()
+
+    async def process_input_frames(self) -> None:
+        while not self.thread_quit.is_set():
+            try:
+                frame = cast(AudioFrame, await self.track.recv())
+                for frame in self.event_handler.resample(frame):
+                    numpy_array = frame.to_ndarray()
+                    if isinstance(self.event_handler, AsyncHandler):
+                        await self.event_handler.receive(
+                            (frame.sample_rate, numpy_array)  # type: ignore
+                        )
+                    else:
+                        await anyio.to_thread.run_sync(
+                            self.event_handler_receive, (frame.sample_rate, numpy_array)
+                        )
+            except MediaStreamError:
+                logger.debug("MediaStreamError in process_input_frames")
+                break
+
+    async def start(self):
+        if not self.has_started:
+            loop = asyncio.get_running_loop()
+            await self.wait_for_channel()
+            if isinstance(self.event_handler, AsyncHandler):
+                callable = self.event_handler.emit
+                start_up = self.event_handler.start_up()
+                if not inspect.isawaitable(start_up):
+                    raise WebRTCError(
+                        "In AsyncStreamHandler, start_up must be a coroutine (async def)"
+                    )
+
+            else:
+                callable = functools.partial(
+                    loop.run_in_executor, None, self.event_handler_emit
+                )
+                start_up = anyio.to_thread.run_sync(self.event_handler.start_up)
+            self.process_input_task = asyncio.create_task(self.process_input_frames())
+            self.process_input_task.add_done_callback(
+                lambda _: logger.debug("process_input_done")
+            )
+            self.start_up_task = asyncio.create_task(start_up)
+            self.start_up_task.add_done_callback(
+                lambda _: logger.debug("start_up_done")
+            )
+            self.decode_task = asyncio.create_task(
+                player_worker_decode(
+                    callable,
+                    self.queue,
+                    self.thread_quit,
+                    lambda: self.channel,
+                    self.set_additional_outputs,
+                    False,
+                    self.event_handler.output_sample_rate,
+                    self.event_handler.output_frame_size,
+                )
+            )
+            self.decode_task.add_done_callback(lambda _: logger.debug("decode_done"))
+            self.has_started = True
+
+    async def recv(self):  # type: ignore
+        try:
+            if self.readyState != "live":
+                raise MediaStreamError
+
+            if not self.event_handler.channel_set.is_set():
+                await self.event_handler.channel_set.wait()
+            if current_channel.get() != self.event_handler.channel:
+                current_channel.set(self.event_handler.channel)
+            await self.start()
+
+            frame = await self.queue.get()
+            logger.debug("frame %s", frame)
+
+            data_time = frame.time
+
+            if time.time() - self.last_timestamp > 10 * (
+                self.event_handler.output_frame_size
+                / self.event_handler.output_sample_rate
+            ):
+                self._start = None
+
+            # control playback rate
+            if self._start is None:
+                self._start = time.time() - data_time  # type: ignore
+            else:
+                wait = self._start + data_time - time.time()
+                await asyncio.sleep(wait)
+            self.last_timestamp = time.time()
+            return frame
+        except Exception as e:
+            logger.debug("exception %s", e)
+            exec = traceback.format_exc()
+            logger.debug("traceback %s", exec)
+
+    def stop(self):
+        logger.debug("audio callback stop")
+        self.thread_quit.set()
+        super().stop()
+
+
+class ServerToClientVideo(VideoStreamTrack):
+    """
+    This works for streaming input and output
+    """
+
+    kind = "video"
+
+    def __init__(
+        self,
+        event_handler: Callable,
+        channel: DataChannel | None = None,
+        set_additional_outputs: Callable | None = None,
+    ) -> None:
+        super().__init__()  # don't forget this!
+        self.event_handler = event_handler
+        self.args_set = asyncio.Event()
+        self.latest_args: str | list[Any] = "not_set"
+        self.generator: Generator[Any, None, Any] | None = None
+        self.channel = channel
+        self.set_additional_outputs = set_additional_outputs
+
+    def array_to_frame(self, array: np.ndarray) -> VideoFrame:
+        return VideoFrame.from_ndarray(array, format="bgr24")
+
+    def set_channel(self, channel: DataChannel):
+        self.channel = channel
+
+    def set_args(self, args: list[Any]):
+        self.latest_args = list(args)
+        self.args_set.set()
+
+    async def recv(self):  # type: ignore
+        try:
+            pts, time_base = await self.next_timestamp()
+            await self.args_set.wait()
+            current_channel.set(self.channel)
+            if self.generator is None:
+                self.generator = cast(
+                    Generator[Any, None, Any], self.event_handler(*self.latest_args)
+                )
+            try:
+                next_array, outputs = split_output(next(self.generator))
+                if (
+                    isinstance(outputs, AdditionalOutputs)
+                    and self.set_additional_outputs
+                    and self.channel
+                ):
+                    self.set_additional_outputs(outputs)
+                    self.channel.send(create_message("fetch_output", []))
+            except StopIteration:
+                self.stop()
+                return
+
+            next_frame = self.array_to_frame(next_array)
+            next_frame.pts = pts
+            next_frame.time_base = time_base
+            return next_frame
+        except Exception as e:
+            logger.debug("exception %s", e)
+            exec = traceback.format_exc()
+            logger.debug("traceback %s %s", e, exec)
+            if isinstance(e, WebRTCError):
+                raise e
+            else:
+                raise WebRTCError(str(e)) from e
+
+
+class ServerToClientAudio(AudioStreamTrack):
+    kind = "audio"
+
+    def __init__(
+        self,
+        event_handler: Callable,
+        channel: DataChannel | None = None,
+        set_additional_outputs: Callable | None = None,
+    ) -> None:
+        self.generator: Generator[Any, None, Any] | None = None
+        self.event_handler = event_handler
+        self.event_handler._clear_queue = self.clear_queue
+        self.current_timestamp = 0
+        self.latest_args: str | list[Any] = "not_set"
+        self.args_set = threading.Event()
+        self.queue = asyncio.Queue()
+        self.thread_quit = asyncio.Event()
+        self.channel = channel
+        self.set_additional_outputs = set_additional_outputs
+        self.has_started = False
+        self._start: float | None = None
+        super().__init__()
+
+    def clear_queue(self):
+        while not self.queue.empty():
+            self.queue.get_nowait()
+        self._start = None
+
+    def set_channel(self, channel: DataChannel):
+        self.channel = channel
+
+    def set_args(self, args: list[Any]):
+        self.latest_args = list(args)
+        self.args_set.set()
+
+    def next(self) -> tuple[int, np.ndarray] | None:
+        self.args_set.wait()
+        current_channel.set(self.channel)
+        if self.generator is None:
+            self.generator = self.event_handler(*self.latest_args)
+        if self.generator is not None:
+            try:
+                frame = next(self.generator)
+                return frame
+            except StopIteration:
+                self.thread_quit.set()
+
+    async def start(self):
+        if not self.has_started:
+            loop = asyncio.get_running_loop()
+            callable = functools.partial(loop.run_in_executor, None, self.next)
+            asyncio.create_task(
+                player_worker_decode(
+                    callable,
+                    self.queue,
+                    self.thread_quit,
+                    lambda: self.channel,
+                    self.set_additional_outputs,
+                    True,
+                )
+            )
+            self.has_started = True
+
+    async def recv(self):  # type: ignore
+        try:
+            if self.readyState != "live":
+                raise MediaStreamError
+
+            await self.start()
+            data = await self.queue.get()
+            if data is None:
+                self.stop()
+                return
+
+            data_time = data.time
+
+            # control playback rate
+            if data_time is not None:
+                if self._start is None:
+                    self._start = time.time() - data_time  # type: ignore
+                else:
+                    wait = self._start + data_time - time.time()
+                    await asyncio.sleep(wait)
+
+            return data
+        except Exception as e:
+            logger.debug("exception %s", e)
+            exec = traceback.format_exc()
+            logger.debug("traceback %s", exec)
+            if isinstance(e, WebRTCError):
+                raise e
+            else:
+                raise WebRTCError(str(e)) from e
+
+    def stop(self):
+        logger.debug("audio-to-client stop callback")
+        self.thread_quit.set()
+        super().stop()

backend/fastrtc/utils.py

@@ -0,0 +1,456 @@
+import asyncio
+import fractions
+import functools
+import inspect
+import io
+import json
+import logging
+import tempfile
+import traceback
+from contextvars import ContextVar
+from typing import Any, Callable, Literal, Protocol, TypedDict, cast
+
+import av
+import numpy as np
+from numpy.typing import NDArray
+from pydub import AudioSegment
+
+logger = logging.getLogger(__name__)
+
+
+AUDIO_PTIME = 0.020
+
+
+class Message(TypedDict):
+    type: str
+    data: Any
+    
+class AudioChunk(TypedDict):
+    start: int
+    end: int
+
+
+class AdditionalOutputs:
+    def __init__(self, *args) -> None:
+        self.args = args
+
+
+class DataChannel(Protocol):
+    def send(self, message: str) -> None: ...
+
+
+def create_message(
+    type: Literal[
+        "send_input",
+        "fetch_output",
+        "stopword",
+        "error",
+        "warning",
+        "log",
+    ],
+    data: list[Any] | str,
+) -> str:
+    return json.dumps({"type": type, "data": data})
+
+
+current_channel: ContextVar[DataChannel | None] = ContextVar(
+    "current_channel", default=None
+)
+
+
+def _send_log(message: str, type: str) -> None:
+    async def _send(channel: DataChannel) -> None:
+        channel.send(
+            json.dumps(
+                {
+                    "type": type,
+                    "message": message,
+                }
+            )
+        )
+
+    if channel := current_channel.get():
+        try:
+            loop = asyncio.get_running_loop()
+            asyncio.run_coroutine_threadsafe(_send(channel), loop)
+        except RuntimeError:
+            asyncio.run(_send(channel))
+
+
+def Warning(  # noqa: N802
+    message: str = "Warning issued.",
+):
+    """
+    Send a warning message that is deplayed in the UI of the application.
+
+    Parameters
+    ----------
+    audio : str
+        The warning message to send
+
+    Returns
+    -------
+    None
+    """
+    _send_log(message, "warning")
+
+
+class WebRTCError(Exception):
+    def __init__(self, message: str) -> None:
+        super().__init__(message)
+        _send_log(message, "error")
+
+
+def split_output(data: tuple | Any) -> tuple[Any, AdditionalOutputs | None]:
+    if isinstance(data, AdditionalOutputs):
+        return None, data
+    if isinstance(data, tuple):
+        # handle the bare audio case
+        if 2 <= len(data) <= 3 and isinstance(data[1], np.ndarray):
+            return data, None
+        if not len(data) == 2:
+            raise ValueError(
+                "The tuple must have exactly two elements: the data and an instance of AdditionalOutputs."
+            )
+        if not isinstance(data[-1], AdditionalOutputs):
+            raise ValueError(
+                "The last element of the tuple must be an instance of AdditionalOutputs."
+            )
+        return data[0], cast(AdditionalOutputs, data[1])
+    return data, None
+
+
+async def player_worker_decode(
+    next_frame: Callable,
+    queue: asyncio.Queue,
+    thread_quit: asyncio.Event,
+    channel: Callable[[], DataChannel | None] | None,
+    set_additional_outputs: Callable | None,
+    quit_on_none: bool = False,
+    sample_rate: int = 48000,
+    frame_size: int = int(48000 * AUDIO_PTIME),
+):
+    audio_samples = 0
+    audio_time_base = fractions.Fraction(1, sample_rate)
+    audio_resampler = av.AudioResampler(  # type: ignore
+        format="s16",
+        layout="stereo",
+        rate=sample_rate,
+        frame_size=frame_size,
+    )
+
+    while not thread_quit.is_set():
+        try:
+            # Get next frame
+            frame, outputs = split_output(
+                await asyncio.wait_for(next_frame(), timeout=60)
+            )
+            if (
+                isinstance(outputs, AdditionalOutputs)
+                and set_additional_outputs
+                and channel
+                and channel()
+            ):
+                set_additional_outputs(outputs)
+                cast(DataChannel, channel()).send(create_message("fetch_output", []))
+
+            if frame is None:
+                if quit_on_none:
+                    await queue.put(None)
+                    break
+                continue
+
+            if not isinstance(frame, tuple) and not isinstance(frame[1], np.ndarray):
+                raise WebRTCError(
+                    "The frame must be a tuple containing a sample rate and a numpy array."
+                )
+
+            if len(frame) == 2:
+                sample_rate, audio_array = frame
+                layout = "mono"
+            elif len(frame) == 3:
+                sample_rate, audio_array, layout = frame
+
+            logger.debug(
+                "received array with shape %s sample rate %s layout %s",
+                audio_array.shape,  # type: ignore
+                sample_rate,
+                layout,  # type: ignore
+            )
+            format = "s16" if audio_array.dtype == "int16" else "fltp"  # type: ignore
+
+            if audio_array.ndim == 1:
+                audio_array = audio_array.reshape(1, -1)
+
+            # Convert to audio frame and resample
+            # This runs in the same timeout context
+            frame = av.AudioFrame.from_ndarray(  # type: ignore
+                audio_array,  # type: ignore
+                format=format,
+                layout=layout,  # type: ignore
+            )
+            frame.sample_rate = sample_rate
+
+            for processed_frame in audio_resampler.resample(frame):
+                processed_frame.pts = audio_samples
+                processed_frame.time_base = audio_time_base
+                audio_samples += processed_frame.samples
+                await queue.put(processed_frame)
+
+        except (TimeoutError, asyncio.TimeoutError):
+            logger.warning(
+                "Timeout in frame processing cycle after %s seconds - resetting", 60
+            )
+            continue
+        except Exception as e:
+            import traceback
+
+            exec = traceback.format_exc()
+            print("traceback %s", exec)
+            print("Error processing frame: %s", str(e))
+            if isinstance(e, WebRTCError):
+                raise e
+            else:
+                continue
+
+
+def audio_to_bytes(audio: tuple[int, NDArray[np.int16 | np.float32]]) -> bytes:
+    """
+    Convert an audio tuple containing sample rate and numpy array data into bytes.
+
+    Parameters
+    ----------
+    audio : tuple[int, np.ndarray]
+        A tuple containing:
+            - sample_rate (int): The audio sample rate in Hz
+            - data (np.ndarray): The audio data as a numpy array
+
+    Returns
+    -------
+    bytes
+        The audio data encoded as bytes, suitable for transmission or storage
+
+    Example
+    -------
+    >>> sample_rate = 44100
+    >>> audio_data = np.array([0.1, -0.2, 0.3])  # Example audio samples
+    >>> audio_tuple = (sample_rate, audio_data)
+    >>> audio_bytes = audio_to_bytes(audio_tuple)
+    """
+    audio_buffer = io.BytesIO()
+    segment = AudioSegment(
+        audio[1].tobytes(),
+        frame_rate=audio[0],
+        sample_width=audio[1].dtype.itemsize,
+        channels=1,
+    )
+    segment.export(audio_buffer, format="mp3")
+    return audio_buffer.getvalue()
+
+
+def audio_to_file(audio: tuple[int, NDArray[np.int16 | np.float32]]) -> str:
+    """
+    Save an audio tuple containing sample rate and numpy array data to a file.
+
+    Parameters
+    ----------
+    audio : tuple[int, np.ndarray]
+        A tuple containing:
+            - sample_rate (int): The audio sample rate in Hz
+            - data (np.ndarray): The audio data as a numpy array
+
+    Returns
+    -------
+    str
+        The path to the saved audio file
+
+    Example
+    -------
+    >>> sample_rate = 44100
+    >>> audio_data = np.array([0.1, -0.2, 0.3])  # Example audio samples
+    >>> audio_tuple = (sample_rate, audio_data)
+    >>> file_path = audio_to_file(audio_tuple)
+    >>> print(f"Audio saved to: {file_path}")
+    """
+    bytes_ = audio_to_bytes(audio)
+    with tempfile.NamedTemporaryFile(suffix=".mp3", delete=False) as f:
+        f.write(bytes_)
+    return f.name
+
+
+def audio_to_float32(
+    audio: tuple[int, NDArray[np.int16 | np.float32]],
+) -> NDArray[np.float32]:
+    """
+    Convert an audio tuple containing sample rate (int16) and numpy array data to float32.
+
+    Parameters
+    ----------
+    audio : tuple[int, np.ndarray]
+        A tuple containing:
+            - sample_rate (int): The audio sample rate in Hz
+            - data (np.ndarray): The audio data as a numpy array
+
+    Returns
+    -------
+    np.ndarray
+        The audio data as a numpy array with dtype float32
+
+    Example
+    -------
+    >>> sample_rate = 44100
+    >>> audio_data = np.array([0.1, -0.2, 0.3])  # Example audio samples
+    >>> audio_tuple = (sample_rate, audio_data)
+    >>> audio_float32 = audio_to_float32(audio_tuple)
+    """
+    return audio[1].astype(np.float32) / 32768.0
+
+
+def audio_to_int16(
+    audio: tuple[int, NDArray[np.int16 | np.float32]],
+) -> NDArray[np.int16]:
+    """
+    Convert an audio tuple containing sample rate and numpy array data to int16.
+
+    Parameters
+    ----------
+    audio : tuple[int, np.ndarray]
+        A tuple containing:
+            - sample_rate (int): The audio sample rate in Hz
+            - data (np.ndarray): The audio data as a numpy array
+
+    Returns
+    -------
+    np.ndarray
+        The audio data as a numpy array with dtype int16
+
+    Example
+    -------
+    >>> sample_rate = 44100
+    >>> audio_data = np.array([0.1, -0.2, 0.3], dtype=np.float32)  # Example audio samples
+    >>> audio_tuple = (sample_rate, audio_data)
+    >>> audio_int16 = audio_to_int16(audio_tuple)
+    """
+    if audio[1].dtype == np.int16:
+        return audio[1]  # type: ignore
+    elif audio[1].dtype == np.float32:
+        # Convert float32 to int16 by scaling to the int16 range
+        return (audio[1] * 32767.0).astype(np.int16)
+    else:
+        raise TypeError(f"Unsupported audio data type: {audio[1].dtype}")
+
+
+def aggregate_bytes_to_16bit(chunks_iterator):
+    """
+    Aggregate bytes to 16-bit audio samples.
+
+    This function takes an iterator of chunks and aggregates them into 16-bit audio samples.
+    It handles incomplete samples and combines them with the next chunk.
+
+    Parameters
+    ----------
+    chunks_iterator : Iterator[bytes]
+        An iterator of byte chunks to aggregate
+
+    Returns
+    -------
+    Iterator[NDArray[np.int16]]
+    """
+    leftover = b""
+    for chunk in chunks_iterator:
+        current_bytes = leftover + chunk
+
+        n_complete_samples = len(current_bytes) // 2
+        bytes_to_process = n_complete_samples * 2
+
+        to_process = current_bytes[:bytes_to_process]
+        leftover = current_bytes[bytes_to_process:]
+
+        if to_process:
+            audio_array = np.frombuffer(to_process, dtype=np.int16).reshape(1, -1)
+            yield audio_array
+
+
+async def async_aggregate_bytes_to_16bit(chunks_iterator):
+    """
+    Aggregate bytes to 16-bit audio samples.
+
+    This function takes an iterator of chunks and aggregates them into 16-bit audio samples.
+    It handles incomplete samples and combines them with the next chunk.
+
+    Parameters
+    ----------
+    chunks_iterator : Iterator[bytes]
+        An iterator of byte chunks to aggregate
+
+    Returns
+    -------
+    Iterator[NDArray[np.int16]]
+        An iterator of 16-bit audio samples
+    """
+    leftover = b""
+
+    async for chunk in chunks_iterator:
+        current_bytes = leftover + chunk
+
+        n_complete_samples = len(current_bytes) // 2
+        bytes_to_process = n_complete_samples * 2
+
+        to_process = current_bytes[:bytes_to_process]
+        leftover = current_bytes[bytes_to_process:]
+
+        if to_process:
+            audio_array = np.frombuffer(to_process, dtype=np.int16).reshape(1, -1)
+            yield audio_array
+
+
+def webrtc_error_handler(func):
+    """Decorator to catch exceptions and raise WebRTCError with stacktrace."""
+
+    @functools.wraps(func)
+    async def async_wrapper(*args, **kwargs):
+        try:
+            return await func(*args, **kwargs)
+        except Exception as e:
+            traceback.print_exc()
+            if isinstance(e, WebRTCError):
+                raise e
+            else:
+                raise WebRTCError(str(e)) from e
+
+    @functools.wraps(func)
+    def sync_wrapper(*args, **kwargs):
+        try:
+            return func(*args, **kwargs)
+        except Exception as e:
+            traceback.print_exc()
+            if isinstance(e, WebRTCError):
+                raise e
+            else:
+                raise WebRTCError(str(e)) from e
+
+    return async_wrapper if inspect.iscoroutinefunction(func) else sync_wrapper
+
+
+async def wait_for_item(queue: asyncio.Queue, timeout: float = 0.1) -> Any:
+    """
+    Wait for an item from an asyncio.Queue with a timeout.
+
+    This function attempts to retrieve an item from the queue using asyncio.wait_for.
+    If the timeout is reached, it returns None.
+
+    This is useful to avoid blocking `emit` when the queue is empty.
+    """
+
+    try:
+        return await asyncio.wait_for(queue.get(), timeout=timeout)
+    except (TimeoutError, asyncio.TimeoutError):
+        return None
+
+def parse_json_safely(str: str):
+    try:
+        result = json.loads(str)
+        return result, None
+    except json.JSONDecodeError as e:
+        print(f"JSON解析错误: {e.msg}")
+        return None, e

backend/fastrtc/webrtc.py

@@ -0,0 +1,370 @@
+"""gr.WebRTC() component."""
+
+from __future__ import annotations
+
+import logging
+# logging.basicConfig(level=logging.DEBUG)
+from collections.abc import Callable
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Concatenate,
+    Iterable,
+    Literal,
+    ParamSpec,
+    Sequence,
+    TypeVar,
+    cast,
+)
+
+from gradio import wasm_utils
+from gradio.components.base import Component, server
+from gradio_client import handle_file
+
+from .tracks import (
+    AudioVideoStreamHandlerImpl,
+    StreamHandler,
+    StreamHandlerBase,
+    StreamHandlerImpl,
+    VideoEventHandler,
+)
+from .webrtc_connection_mixin import WebRTCConnectionMixin
+
+if TYPE_CHECKING:
+    from gradio.blocks import Block
+    from gradio.components import Timer
+
+if wasm_utils.IS_WASM:
+    raise ValueError("Not supported in gradio-lite!")
+
+
+logger = logging.getLogger(__name__)
+
+
+# For the return type
+R = TypeVar("R")
+# For the parameter specification
+P = ParamSpec("P")
+
+
+class WebRTC(Component, WebRTCConnectionMixin):
+    """
+    Creates a video component that can be used to upload/record videos (as an input) or display videos (as an output).
+    For the video to be playable in the browser it must have a compatible container and codec combination. Allowed
+    combinations are .mp4 with h264 codec, .ogg with theora codec, and .webm with vp9 codec. If the component detects
+    that the output video would not be playable in the browser it will attempt to convert it to a playable mp4 video.
+    If the conversion fails, the original video is returned.
+
+    Demos: video_identity_2
+    """
+
+    EVENTS = ["tick", "state_change"]
+
+    def __init__(
+        self,
+        value: None = None,
+        height: int | str | None = None,
+        width: int | str | None = None,
+        label: str | None = None,
+        every: Timer | float | None = None,
+        inputs: Component | Sequence[Component] | set[Component] | None = None,
+        show_label: bool | None = None,
+        container: bool = True,
+        scale: int | None = None,
+        min_width: int = 160,
+        interactive: bool | None = None,
+        visible: bool = True,
+        elem_id: str | None = None,
+        elem_classes: list[str] | str | None = None,
+        render: bool = True,
+        key: int | str | None = None,
+        mirror_webcam: bool = True,
+        rtc_configuration: dict[str, Any] | None = None,
+        track_constraints: dict[str, Any] | None = None,
+        time_limit: float | None = None,
+        mode: Literal["send-receive", "receive", "send"] = "send-receive",
+        modality: Literal["video", "audio", "audio-video"] = "video",
+        rtp_params: dict[str, Any] | None = None,
+        icon: str | None = None,
+        icon_button_color: str | None = None,
+        pulse_color: str | None = None,
+        icon_radius: int | None = None,
+        button_labels: dict | None = None,
+        video_chat: bool = True,
+    ):
+        """
+        Parameters:
+            value: path or URL for the default value that WebRTC component is going to take. Can also be a tuple consisting of (video filepath, subtitle filepath). If a subtitle file is provided, it should be of type .srt or .vtt. Or can be callable, in which case the function will be called whenever the app loads to set the initial value of the component.
+            format: the file extension with which to save video, such as 'avi' or 'mp4'. This parameter applies both when this component is used as an input to determine which file format to convert user-provided video to, and when this component is used as an output to determine the format of video returned to the user. If None, no file format conversion is done and the video is kept as is. Use 'mp4' to ensure browser playability.
+            height: The height of the component, specified in pixels if a number is passed, or in CSS units if a string is passed. This has no effect on the preprocessed video file, but will affect the displayed video.
+            width: The width of the component, specified in pixels if a number is passed, or in CSS units if a string is passed. This has no effect on the preprocessed video file, but will affect the displayed video.
+            label: the label for this component. Appears above the component and is also used as the header if there are a table of examples for this component. If None and used in a `gr.Interface`, the label will be the name of the parameter this component is assigned to.
+            every: continously calls `value` to recalculate it if `value` is a function (has no effect otherwise). Can provide a Timer whose tick resets `value`, or a float that provides the regular interval for the reset Timer.
+            inputs: components that are used as inputs to calculate `value` if `value` is a function (has no effect otherwise). `value` is recalculated any time the inputs change.
+            show_label: if True, will display label.
+            container: if True, will place the component in a container - providing some extra padding around the border.
+            scale: relative size compared to adjacent Components. For example if Components A and B are in a Row, and A has scale=2, and B has scale=1, A will be twice as wide as B. Should be an integer. scale applies in Rows, and to top-level Components in Blocks where fill_height=True.
+            min_width: minimum pixel width, will wrap if not sufficient screen space to satisfy this value. If a certain scale value results in this Component being narrower than min_width, the min_width parameter will be respected first.
+            interactive: if True, will allow users to upload a video; if False, can only be used to display videos. If not provided, this is inferred based on whether the component is used as an input or output.
+            visible: if False, component will be hidden.
+            elem_id: an optional string that is assigned as the id of this component in the HTML DOM. Can be used for targeting CSS styles.
+            elem_classes: an optional list of strings that are assigned as the classes of this component in the HTML DOM. Can be used for targeting CSS styles.
+            render: if False, component will not render be rendered in the Blocks context. Should be used if the intention is to assign event listeners now but render the component later.
+            key: if assigned, will be used to assume identity across a re-render. Components that have the same key across a re-render will have their value preserved.
+            mirror_webcam: if True webcam will be mirrored. Default is True.
+            rtc_configuration: WebRTC configuration options. See https://developer.mozilla.org/en-US/docs/Web/API/RTCPeerConnection/RTCPeerConnection . If running the demo on a remote server, you will need to specify a rtc_configuration. See https://freddyaboulton.github.io/gradio-webrtc/deployment/
+            track_constraints: Media track constraints for WebRTC. For example, to set video height, width use {"width": {"exact": 800}, "height": {"exact": 600}, "aspectRatio": {"exact": 1.33333}}
+            time_limit: Maximum duration in seconds for recording.
+            mode: WebRTC mode - "send-receive", "receive", or "send".
+            modality: Type of media - "video" or "audio".
+            rtp_params: See https://developer.mozilla.org/en-US/docs/Web/API/RTCRtpSender/setParameters. If you are changing the video resolution, you can set this to {"degradationPreference": "maintain-framerate"} to keep the frame rate consistent.
+            icon: Icon to display on the button instead of the wave animation. The icon should be a path/url to a .svg/.png/.jpeg file.
+            icon_button_color: Color of the icon button. Default is var(--color-accent) of the demo theme.
+            pulse_color: Color of the pulse animation. Default is var(--color-accent) of the demo theme.
+            button_labels: Text to display on the audio or video start, stop, waiting buttons. Dict with keys "start", "stop", "waiting" mapping to the text to display on the buttons.
+            icon_radius: Border radius of the icon button expressed as a percentage of the button size. Default is 50%
+        """
+        WebRTCConnectionMixin.__init__(self)
+        self.time_limit = time_limit
+        self.height = height
+        self.width = width
+        self.mirror_webcam = mirror_webcam
+        self.concurrency_limit = 1
+        self.rtc_configuration = rtc_configuration
+        self.mode = mode
+        self.modality = modality
+        self.icon_button_color = icon_button_color
+        self.icon_radius = icon_radius
+        self.pulse_color = pulse_color
+        self.rtp_params = rtp_params or {}
+        self.video_chat = video_chat
+        self.button_labels = {
+            "start": "",
+            "stop": "",
+            "waiting": "",
+            **(button_labels or {}),
+        }
+        if track_constraints is None and modality == "audio":
+            track_constraints = {
+                "echoCancellation": True,
+                "noiseSuppression": {"exact": True},
+                "autoGainControl": {"exact": True},
+                "sampleRate": {"ideal": 24000},
+                "sampleSize": {"ideal": 16},
+                "channelCount": {"exact": 1},
+            }
+        if track_constraints is None and modality == "video":
+            track_constraints = {
+                "facingMode": "user",
+                "width": {"ideal": 500},
+                "height": {"ideal": 500},
+                "frameRate": {"ideal": 30},
+            }
+        if track_constraints is None and modality == "audio-video":
+            track_constraints = {
+                "video": {
+                    "facingMode": "user",
+                    "width": {"ideal": 500},
+                    "height": {"ideal": 500},
+                    "frameRate": {"ideal": 30},
+                },
+                "audio": {
+                    "echoCancellation": True,
+                    "noiseSuppression": {"exact": True},
+                    "autoGainControl": {"exact": True},
+                    "sampleRate": {"ideal": 24000},
+                    "sampleSize": {"ideal": 16},
+                    "channelCount": {"exact": 1},
+                },
+            }
+        self.track_constraints = track_constraints
+        self.event_handler: Callable | StreamHandler | None = None
+        super().__init__(
+            label=label,
+            every=every,
+            inputs=inputs,
+            show_label=show_label,
+            container=container,
+            scale=scale,
+            min_width=min_width,
+            interactive=interactive,
+            visible=visible,
+            elem_id=elem_id,
+            elem_classes=elem_classes,
+            render=render,
+            key=key,
+            value=value,
+        )
+        # need to do this here otherwise the proxy_url is not set
+        self.icon = (
+            icon if not icon else cast(dict, self.serve_static_file(icon)).get("url")
+        )
+
+    def preprocess(self, payload: str) -> str:
+        """
+        Parameters:
+            payload: An instance of VideoData containing the video and subtitle files.
+        Returns:
+            Passes the uploaded video as a `str` filepath or URL whose extension can be modified by `format`.
+        """
+        return payload
+
+    def postprocess(self, value: Any) -> str:
+        """
+        Parameters:
+            value: Expects a {str} or {pathlib.Path} filepath to a video which is displayed, or a {Tuple[str | pathlib.Path, str | pathlib.Path | None]} where the first element is a filepath to a video and the second element is an optional filepath to a subtitle file.
+        Returns:
+            VideoData object containing the video and subtitle files.
+        """
+        return value
+
+    def on_additional_outputs(
+        self,
+        fn: Callable[Concatenate[P], R],
+        inputs: Block | Sequence[Block] | set[Block] | None = None,
+        outputs: Block | Sequence[Block] | set[Block] | None = None,
+        js: str | None = None,
+        concurrency_limit: int | None | Literal["default"] = "default",
+        concurrency_id: str | None = None,
+        show_progress: Literal["full", "minimal", "hidden"] = "full",
+        queue: bool = True,
+    ):
+        inputs = inputs or []
+        if inputs and not isinstance(inputs, Iterable):
+            inputs = [inputs]
+            inputs = list(inputs)
+
+        async def handler(webrtc_id: str, *args):
+            async for next_outputs in self.output_stream(webrtc_id):
+                yield fn(*args, *next_outputs.args)  # type: ignore
+
+        return self.state_change(  # type: ignore
+            fn=handler,
+            inputs=[self] + cast(list, inputs),
+            outputs=outputs,
+            js=js,
+            concurrency_limit=concurrency_limit,
+            concurrency_id=concurrency_id,
+            show_progress="minimal",
+            queue=queue,
+            trigger_mode="once",
+        )
+
+    def stream(
+        self,
+        fn: (
+            Callable[..., Any]
+            | StreamHandlerImpl
+            | AudioVideoStreamHandlerImpl
+            | VideoEventHandler
+            | None
+        ) = None,
+        inputs: Block | Sequence[Block] | set[Block] | None = None,
+        outputs: Block | Sequence[Block] | set[Block] | None = None,
+        js: str | None = None,
+        concurrency_limit: int | None | Literal["default"] = "default",
+        concurrency_id: str | None = None,
+        time_limit: float | None = None,
+        trigger: Callable | None = None,
+    ):
+        from gradio.blocks import Block
+
+        if inputs is None:
+            inputs = []
+        if outputs is None:
+            outputs = []
+        if isinstance(inputs, Block):
+            inputs = [inputs]
+        if isinstance(outputs, Block):
+            outputs = [outputs]
+
+        self.concurrency_limit = cast(
+            int, (1 if concurrency_limit in ["default", None] else concurrency_limit)
+        )
+        self.event_handler = fn  # type: ignore
+        self.time_limit = time_limit
+
+        if (
+            self.mode == "send-receive"
+            and self.modality in ["audio", "audio-video"]
+            and not isinstance(self.event_handler, StreamHandlerBase)
+        ):
+            raise ValueError(
+                "In the send-receive mode for audio, the event handler must be an instance of StreamHandlerBase."
+            )
+
+        if self.mode == "send-receive" or self.mode == "send":
+            if cast(list[Block], inputs)[0] != self:
+                raise ValueError(
+                    "In the webrtc stream event, the first input component must be the WebRTC component."
+                )
+
+            if (
+                len(cast(list[Block], outputs)) != 1
+                and cast(list[Block], outputs)[0] != self
+            ):
+                raise ValueError(
+                    "In the webrtc stream event, the only output component must be the WebRTC component."
+                )
+            for input_component in inputs[1:]:  # type: ignore
+                if hasattr(input_component, "change"):
+                    input_component.change(  # type: ignore
+                        self.set_input,
+                        inputs=inputs,
+                        outputs=None,
+                        concurrency_id=concurrency_id,
+                        concurrency_limit=None,
+                        time_limit=None,
+                        js=js,
+                    )
+            return self.tick(  # type: ignore
+                self.set_input,
+                inputs=inputs,
+                outputs=None,
+                concurrency_id=concurrency_id,
+                concurrency_limit=None,
+                time_limit=None,
+                js=js,
+            )
+        elif self.mode == "receive":
+            if isinstance(inputs, list) and self in cast(list[Block], inputs):
+                raise ValueError(
+                    "In the receive mode stream event, the WebRTC component cannot be an input."
+                )
+            if (
+                len(cast(list[Block], outputs)) != 1
+                and cast(list[Block], outputs)[0] != self
+            ):
+                raise ValueError(
+                    "In the receive mode stream, the only output component must be the WebRTC component."
+                )
+            if trigger is None:
+                raise ValueError(
+                    "In the receive mode stream event, the trigger parameter must be provided"
+                )
+            trigger(lambda: "start_webrtc_stream", inputs=None, outputs=self)
+            self.tick(  # type: ignore
+                self.set_input,
+                inputs=[self] + list(inputs),
+                outputs=None,
+                concurrency_id=concurrency_id,
+            )
+
+    @server
+    async def offer(self, body):
+        return await self.handle_offer(
+            body, self.set_additional_outputs(body["webrtc_id"])
+        )
+
+    def example_payload(self) -> Any:
+        return {
+            "video": handle_file(
+                "https://github.com/gradio-app/gradio/raw/main/demo/video_component/files/world.mp4"
+            ),
+        }
+
+    def example_value(self) -> Any:
+        return "https://github.com/gradio-app/gradio/raw/main/demo/video_component/files/world.mp4"
+
+    def api_info(self) -> Any:
+        return {"type": "number"}

backend/fastrtc/webrtc_connection_mixin.py

@@ -0,0 +1,311 @@
+"""Mixin for handling WebRTC connections."""
+
+from __future__ import annotations
+
+import json
+import asyncio
+import inspect
+import logging
+from collections import defaultdict
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from typing import (
+    AsyncGenerator,
+    Literal,
+    ParamSpec,
+    TypeVar,
+    cast,
+)
+
+from aiortc import (
+    RTCPeerConnection,
+    RTCSessionDescription,
+)
+from aiortc.contrib.media import MediaRelay  # type: ignore
+from fastapi.responses import JSONResponse
+
+from fastrtc.tracks import (
+    AudioCallback,
+    HandlerType,
+    ServerToClientAudio,
+    ServerToClientVideo,
+    StreamHandlerBase,
+    StreamHandlerImpl,
+    VideoCallback,
+    VideoStreamHandler,
+)
+from fastrtc.utils import (
+    AdditionalOutputs,
+    Message,
+    create_message,
+    parse_json_safely,
+    webrtc_error_handler,
+)
+
+Track = (
+    VideoCallback
+    | VideoStreamHandler
+    | AudioCallback
+    | ServerToClientAudio
+    | ServerToClientVideo
+)
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.DEBUG)
+
+
+# For the return type
+R = TypeVar("R")
+# For the parameter specification
+P = ParamSpec("P")
+
+
+@dataclass
+class OutputQueue:
+    queue: asyncio.Queue[AdditionalOutputs] = field(default_factory=asyncio.Queue)
+    quit: asyncio.Event = field(default_factory=asyncio.Event)
+
+
+class WebRTCConnectionMixin:
+    def __init__(self):
+        self.pcs = set([])
+        self.relay = MediaRelay()
+        self.connections = defaultdict(list)
+        self.data_channels = {}
+        self.additional_outputs = defaultdict(OutputQueue)
+        self.handlers = {}
+        self.connection_timeouts = defaultdict(asyncio.Event)
+        # These attributes should be set by subclasses:
+        self.concurrency_limit: int | float | None
+        self.event_handler: HandlerType | None
+        self.time_limit: float | None
+        self.modality: Literal["video", "audio", "audio-video"]
+        self.mode: Literal["send", "receive", "send-receive"]
+
+    @staticmethod
+    async def wait_for_time_limit(pc: RTCPeerConnection, time_limit: float):
+        await asyncio.sleep(time_limit)
+        await pc.close()
+
+    async def connection_timeout(
+        self,
+        pc: RTCPeerConnection,
+        webrtc_id: str,
+        time_limit: float,
+    ):
+        try:
+            await asyncio.wait_for(
+                self.connection_timeouts[webrtc_id].wait(), time_limit
+            )
+        except (asyncio.TimeoutError, TimeoutError):
+            await pc.close()
+            self.connection_timeouts[webrtc_id].clear()
+            self.clean_up(webrtc_id)
+
+    def clean_up(self, webrtc_id: str):
+        self.handlers.pop(webrtc_id, None)
+        self.connection_timeouts.pop(webrtc_id, None)
+        connection = self.connections.pop(webrtc_id, [])
+        for conn in connection:
+            if isinstance(conn, AudioCallback):
+                if inspect.iscoroutinefunction(conn.event_handler.shutdown):
+                    asyncio.create_task(conn.event_handler.shutdown())
+                    conn.event_handler.reset()
+                else:
+                    conn.event_handler.shutdown()
+                    conn.event_handler.reset()
+        output = self.additional_outputs.pop(webrtc_id, None)
+        if output:
+            logger.debug("setting quit for webrtc id %s", webrtc_id)
+            output.quit.set()
+        self.data_channels.pop(webrtc_id, None)
+        return connection
+
+    def set_input(self, webrtc_id: str, *args):
+        if webrtc_id in self.connections:
+            for conn in self.connections[webrtc_id]:
+                conn.set_args(list(args))
+
+    async def output_stream(
+        self, webrtc_id: str
+    ) -> AsyncGenerator[AdditionalOutputs, None]:
+        outputs = self.additional_outputs[webrtc_id]
+        while not outputs.quit.is_set():
+            try:
+                yield await asyncio.wait_for(outputs.queue.get(), 10)
+            except (asyncio.TimeoutError, TimeoutError):
+                logger.debug("Timeout waiting for output")
+
+    async def fetch_latest_output(self, webrtc_id: str) -> AdditionalOutputs:
+        outputs = self.additional_outputs[webrtc_id]
+        return await asyncio.wait_for(outputs.queue.get(), 10)
+
+    def set_additional_outputs(
+        self, webrtc_id: str
+    ) -> Callable[[AdditionalOutputs], None]:
+        def set_outputs(outputs: AdditionalOutputs):
+            self.additional_outputs[webrtc_id].queue.put_nowait(outputs)
+
+        return set_outputs
+
+    async def handle_offer(self, body, set_outputs):
+        logger.debug("Starting to handle offer")
+        logger.debug("Offer body %s", body)
+        if len(self.connections) >= cast(int, self.concurrency_limit):
+            return JSONResponse(
+                status_code=200,
+                content={
+                    "status": "failed",
+                    "meta": {
+                        "error": "concurrency_limit_reached",
+                        "limit": self.concurrency_limit,
+                    },
+                },
+            )
+
+        offer = RTCSessionDescription(sdp=body["sdp"], type=body["type"])
+
+        pc = RTCPeerConnection()
+        self.pcs.add(pc)
+
+        if isinstance(self.event_handler, StreamHandlerBase):
+            handler = self.event_handler.copy()
+            handler.emit = webrtc_error_handler(handler.emit)  # type: ignore
+            handler.receive = webrtc_error_handler(handler.receive)  # type: ignore
+            handler.start_up = webrtc_error_handler(handler.start_up)  # type: ignore
+            handler.shutdown = webrtc_error_handler(handler.shutdown)  # type: ignore
+            if hasattr(handler, "video_receive"):
+                handler.video_receive = webrtc_error_handler(handler.video_receive)  # type: ignore
+            if hasattr(handler, "video_emit"):
+                handler.video_emit = webrtc_error_handler(handler.video_emit)  # type: ignore
+            if hasattr(handler, "on_chat_datachannel"):
+                handler.on_chat_datachannel = webrtc_error_handler(handler.on_chat_datachannel)  # type: ignore
+        else:
+            handler = webrtc_error_handler(cast(Callable, self.event_handler))
+
+        self.handlers[body["webrtc_id"]] = handler
+
+        @pc.on("iceconnectionstatechange")
+        async def on_iceconnectionstatechange():
+            logger.debug("ICE connection state change %s", pc.iceConnectionState)
+            if pc.iceConnectionState == "failed":
+                await pc.close()
+                self.connections.pop(body["webrtc_id"], None)
+                self.pcs.discard(pc)
+
+        @pc.on("connectionstatechange")
+        async def _():
+            print("pc.connectionState %s", pc.connectionState)
+            logger.debug("pc.connectionState %s", pc.connectionState)
+            if pc.connectionState in ["failed", "closed"]:
+                await pc.close()
+                connection = self.clean_up(body["webrtc_id"])
+                if connection:
+                    for conn in connection:
+                        conn.stop()
+                self.pcs.discard(pc)
+            if pc.connectionState == "connected":
+                self.connection_timeouts[body["webrtc_id"]].set()
+                if self.time_limit is not None:
+                    asyncio.create_task(self.wait_for_time_limit(pc, self.time_limit))
+
+        @pc.on("track")
+        def _(track):
+            relay = MediaRelay()
+            handler = self.handlers[body["webrtc_id"]]
+            if self.modality == "video" and track.kind == "video":
+                cb = VideoCallback(
+                    relay.subscribe(track),
+                    event_handler=cast(Callable, handler),
+                    set_additional_outputs=set_outputs,
+                    mode=cast(Literal["send", "send-receive"], self.mode),
+                )
+            elif self.modality == "audio-video" and track.kind == "video":
+                cb = VideoStreamHandler(
+                    relay.subscribe(track),
+                    event_handler=handler,  # type: ignore
+                    set_additional_outputs=set_outputs,
+                )
+            elif self.modality in ["audio", "audio-video"] and track.kind == "audio":
+                eh = cast(StreamHandlerImpl, handler)
+                eh._loop = asyncio.get_running_loop()
+                cb = AudioCallback(
+                    relay.subscribe(track),
+                    event_handler=eh,
+                    set_additional_outputs=set_outputs,
+                )
+            else:
+                raise ValueError("Modality must be either video, audio, or audio-video")
+            if body["webrtc_id"] not in self.connections:
+                self.connections[body["webrtc_id"]] = []
+
+            self.connections[body["webrtc_id"]].append(cb)
+            if body["webrtc_id"] in self.data_channels:
+                for conn in self.connections[body["webrtc_id"]]:
+                    conn.set_channel(self.data_channels[body["webrtc_id"]])
+            if self.mode == "send-receive":
+                logger.debug("Adding track to peer connection %s", cb)
+                pc.addTrack(cb)
+            elif self.mode == "send":
+                asyncio.create_task(cast(AudioCallback | VideoCallback, cb).start())
+
+        if self.mode == "receive":
+            if self.modality == "video":
+                cb = ServerToClientVideo(
+                    cast(Callable, self.event_handler),
+                    set_additional_outputs=set_outputs,
+                )
+            elif self.modality == "audio":
+                cb = ServerToClientAudio(
+                    cast(Callable, self.event_handler),
+                    set_additional_outputs=set_outputs,
+                )
+            else:
+                raise ValueError("Modality must be either video or audio")
+
+            logger.debug("Adding track to peer connection %s", cb)
+            pc.addTrack(cb)
+            self.connections[body["webrtc_id"]].append(cb)
+            cb.on("ended", lambda: self.clean_up(body["webrtc_id"]))
+
+        @pc.on("datachannel")
+        def _(channel):
+            logger.debug(f"Data channel established: {channel.label}")
+
+            self.data_channels[body["webrtc_id"]] = channel
+
+            async def set_channel(webrtc_id: str):
+                while not self.connections.get(webrtc_id):
+                    await asyncio.sleep(0.05)
+                logger.debug("setting channel for webrtc id %s", webrtc_id)
+                for conn in self.connections[webrtc_id]:
+                    conn.set_channel(channel)
+
+            asyncio.create_task(set_channel(body["webrtc_id"]))
+
+            @channel.on("message")
+            def _(message):
+                logger.debug(f"Received message: {message}")
+                if channel.readyState == "open":
+                  msg_dict,error = parse_json_safely(message)
+                  if(error is None and msg_dict['type'] in ['chat','stop_chat']):
+                    msg_dict = cast(Message, json.loads(message))
+                    asyncio.create_task(self.handlers[body["webrtc_id"]].on_chat_datachannel(msg_dict,channel))
+                  else:
+                    channel.send(
+                        create_message("log", data=f"Server received: {message}")
+                    )
+
+        # handle offer
+        await pc.setRemoteDescription(offer)
+        asyncio.create_task(self.connection_timeout(pc, body["webrtc_id"], 30))
+        # send answer
+        answer = await pc.createAnswer()
+        await pc.setLocalDescription(answer)  # type: ignore
+        logger.debug("done handling offer about to return")
+        await asyncio.sleep(0.1)
+
+        return {
+            "sdp": pc.localDescription.sdp,
+            "type": pc.localDescription.type,
+        }

backend/fastrtc/websocket.py

@@ -0,0 +1,215 @@
+import asyncio
+import audioop
+import base64
+import logging
+from typing import Any, Awaitable, Callable, Optional, cast
+
+import anyio
+import librosa
+import numpy as np
+from fastapi import WebSocket
+
+from .tracks import AsyncStreamHandler, StreamHandlerImpl
+from .utils import AdditionalOutputs, DataChannel, split_output
+
+
+class WebSocketDataChannel(DataChannel):
+    def __init__(self, websocket: WebSocket, loop: asyncio.AbstractEventLoop):
+        self.websocket = websocket
+        self.loop = loop
+
+    def send(self, message: str) -> None:
+        asyncio.run_coroutine_threadsafe(self.websocket.send_text(message), self.loop)
+
+
+logger = logging.getLogger(__file__)
+
+
+def convert_to_mulaw(
+    audio_data: np.ndarray, original_rate: int, target_rate: int
+) -> bytes:
+    """Convert audio data to 8kHz mu-law format"""
+
+    if audio_data.dtype != np.float32:
+        audio_data = audio_data.astype(np.float32) / 32768.0
+
+    if original_rate != target_rate:
+        audio_data = librosa.resample(audio_data, orig_sr=original_rate, target_sr=8000)
+
+    audio_data = (audio_data * 32768).astype(np.int16)
+
+    return audioop.lin2ulaw(audio_data, 2)  # type: ignore
+
+
+run_sync = anyio.to_thread.run_sync  # type: ignore
+
+
+class WebSocketHandler:
+    def __init__(
+        self,
+        stream_handler: StreamHandlerImpl,
+        set_handler: Callable[[str, "WebSocketHandler"], Awaitable[None]],
+        clean_up: Callable[[str], None],
+        additional_outputs_factory: Callable[
+            [str], Callable[[AdditionalOutputs], None]
+        ],
+    ):
+        self.stream_handler = stream_handler
+        self.stream_handler._clear_queue = self._clear_queue
+        self.websocket: Optional[WebSocket] = None
+        self._emit_task: Optional[asyncio.Task] = None
+        self.stream_id: Optional[str] = None
+        self.set_additional_outputs_factory = additional_outputs_factory
+        self.set_additional_outputs: Callable[[AdditionalOutputs], None]
+        self.set_handler = set_handler
+        self.quit = asyncio.Event()
+        self.clean_up = clean_up
+        self.queue = asyncio.Queue()
+
+    def _clear_queue(self):
+        old_queue = self.queue
+        self.queue = asyncio.Queue()
+        logger.debug("clearing queue")
+        i = 0
+        while not old_queue.empty():
+            try:
+                old_queue.get_nowait()
+                i += 1
+            except asyncio.QueueEmpty:
+                break
+        logger.debug("popped %d items from queue", i)
+
+    def set_args(self, args: list[Any]):
+        self.stream_handler.set_args(args)
+
+    async def handle_websocket(self, websocket: WebSocket):
+        await websocket.accept()
+        loop = asyncio.get_running_loop()
+        self.loop = loop
+        self.websocket = websocket
+        self.data_channel = WebSocketDataChannel(websocket, loop)
+        self.stream_handler._loop = loop
+        self.stream_handler.set_channel(self.data_channel)
+        self._emit_task = asyncio.create_task(self._emit_loop())
+        self._emit_to_queue_task = asyncio.create_task(self._emit_to_queue())
+        if isinstance(self.stream_handler, AsyncStreamHandler):
+            start_up = self.stream_handler.start_up()
+        else:
+            start_up = anyio.to_thread.run_sync(self.stream_handler.start_up)  # type: ignore
+
+        self.start_up_task = asyncio.create_task(start_up)
+        try:
+            while not self.quit.is_set():
+                message = await websocket.receive_json()
+
+                if message["event"] == "media":
+                    audio_payload = base64.b64decode(message["media"]["payload"])
+
+                    audio_array = np.frombuffer(
+                        audioop.ulaw2lin(audio_payload, 2), dtype=np.int16
+                    )
+
+                    if self.stream_handler.input_sample_rate != 8000:
+                        audio_array = audio_array.astype(np.float32) / 32768.0
+                        audio_array = librosa.resample(
+                            audio_array,
+                            orig_sr=8000,
+                            target_sr=self.stream_handler.input_sample_rate,
+                        )
+                        audio_array = (audio_array * 32768).astype(np.int16)
+                    if isinstance(self.stream_handler, AsyncStreamHandler):
+                        await self.stream_handler.receive(
+                            (self.stream_handler.input_sample_rate, audio_array)
+                        )
+                    else:
+                        await run_sync(
+                            self.stream_handler.receive,
+                            (self.stream_handler.input_sample_rate, audio_array),
+                        )
+
+                elif message["event"] == "start":
+                    if self.stream_handler.phone_mode:
+                        self.stream_id = cast(str, message["streamSid"])
+                    else:
+                        self.stream_id = cast(str, message["websocket_id"])
+                    self.set_additional_outputs = self.set_additional_outputs_factory(
+                        self.stream_id
+                    )
+                    await self.set_handler(self.stream_id, self)
+                elif message["event"] == "stop":
+                    self.quit.set()
+                    self.clean_up(cast(str, self.stream_id))
+                    return
+                elif message["event"] == "ping":
+                    await websocket.send_json({"event": "pong"})
+
+        except Exception as e:
+            print(e)
+            import traceback
+
+            traceback.print_exc()
+            logger.debug("Error in websocket handler %s", e)
+        finally:
+            if self._emit_task:
+                self._emit_task.cancel()
+            if self._emit_to_queue_task:
+                self._emit_to_queue_task.cancel()
+            if self.start_up_task:
+                self.start_up_task.cancel()
+            await websocket.close()
+
+    async def _emit_to_queue(self):
+        try:
+            while not self.quit.is_set():
+                if isinstance(self.stream_handler, AsyncStreamHandler):
+                    output = await self.stream_handler.emit()
+                else:
+                    output = await run_sync(self.stream_handler.emit)
+                self.queue.put_nowait(output)
+        except asyncio.CancelledError:
+            logger.debug("Emit loop cancelled")
+        except Exception as e:
+            import traceback
+
+            traceback.print_exc()
+            logger.debug("Error in emit loop: %s", e)
+
+    async def _emit_loop(self):
+        try:
+            while not self.quit.is_set():
+                output = await self.queue.get()
+
+                if output is not None:
+                    frame, output = split_output(output)
+                    if output is not None:
+                        self.set_additional_outputs(output)
+                    if not isinstance(frame, tuple):
+                        continue
+                    target_rate = (
+                        self.stream_handler.output_sample_rate
+                        if not self.stream_handler.phone_mode
+                        else 8000
+                    )
+                    mulaw_audio = convert_to_mulaw(
+                        frame[1], frame[0], target_rate=target_rate
+                    )
+                    audio_payload = base64.b64encode(mulaw_audio).decode("utf-8")
+
+                    if self.websocket and self.stream_id:
+                        payload = {
+                            "event": "media",
+                            "media": {"payload": audio_payload},
+                        }
+                        if self.stream_handler.phone_mode:
+                            payload["streamSid"] = self.stream_id
+                        await self.websocket.send_json(payload)
+
+                await asyncio.sleep(0.02)
+
+        except asyncio.CancelledError:
+            logger.debug("Emit loop cancelled")
+        except Exception as e:
+            import traceback
+
+            traceback.print_exc()
+            logger.debug("Error in emit loop: %s", e)

demo/app.py

@@ -1,10 +1,17 @@
 import asyncio
 import base64
 from io import BytesIO
+import json
+import math
+import queue
+import time
+import uuid
+import threading
 
+from fastrtc.utils import Message
 import gradio as gr
 import numpy as np
-from gradio_webrtc import (
+from fastrtc import (
     AsyncAudioVideoStreamHandler,
     WebRTC,
     VideoEmitType,
@@ -26,6 +33,7 @@ def encode_image(data: np.ndarray) -> dict:
     base64_str = str(base64.b64encode(bytes_data), "utf-8")
     return {"mime_type": "image/jpeg", "data": base64_str}
 
+frame_queue = queue.Queue(maxsize=100)
 
 class VideoChatHandler(AsyncAudioVideoStreamHandler):
     def __init__(
@@ -38,7 +46,7 @@ class VideoChatHandler(AsyncAudioVideoStreamHandler):
             input_sample_rate=24000,
         )
         self.audio_queue = asyncio.Queue()
-        self.video_queue = asyncio.Queue()
+        self.video_queue = frame_queue
         self.quit = asyncio.Event()
         self.session = None
         self.last_frame_time = 0
@@ -50,6 +58,25 @@ class VideoChatHandler(AsyncAudioVideoStreamHandler):
             output_frame_size=self.output_frame_size,
         )
     
+    chat_id = ''
+    async def on_chat_datachannel(self,message: Message,channel): 
+      # 返回
+      # {"type":"chat",id:"标识属于同一段话", "message":"Hello, world!"}
+      # {"type":"avatar_end"} 表示本次对话结束
+      if message['type'] == 'stop_chat':
+        self.chat_id = ''
+        channel.send(json.dumps({'type':'avatar_end'}))
+      else:
+        id = uuid.uuid4().hex
+        self.chat_id = id
+        data = message["data"]
+        halfLen =  math.floor(data.__len__()/2)
+        channel.send(json.dumps({"type":"chat","id":id,"message":data[:halfLen]}))
+        await asyncio.sleep(5)
+        if self.chat_id == id:
+          channel.send(json.dumps({"type":"chat","id":id,"message":data[halfLen:]}))
+          channel.send(json.dumps({'type':'avatar_end'}))
+    
     async def video_receive(self, frame: np.ndarray):
         # if self.session:
         #     # send image every 1 second
@@ -61,10 +88,11 @@ class VideoChatHandler(AsyncAudioVideoStreamHandler):
         # print(frame.shape)
         newFrame = np.array(frame)
         newFrame[0:, :, 0] = 255 - newFrame[0:, :, 0]
-        self.video_queue.put_nowait(newFrame)
+        # self.video_queue.put_nowait(newFrame)
     
     async def video_emit(self) -> VideoEmitType:
-        return await self.video_queue.get()
+        # print('123123',frame_queue.qsize())
+        return frame_queue.get()
 
     async def receive(self, frame: tuple[int, np.ndarray]) -> None:
         frame_size, array = frame
@@ -114,14 +142,35 @@ with gr.Blocks(css=css) as demo:
                 },
             }
         )
+        handler = VideoChatHandler()
         webrtc.stream(
-            VideoChatHandler(),
+            handler,
             inputs=[webrtc],
             outputs=[webrtc],
-            time_limit=150,
+            time_limit=1500,
             concurrency_limit=2,
         )
-
+        # 线程函数：随机生成 numpy 帧
+        def generate_frames(width=480, height=960, channels=3):
+            while True:
+                try:
+                    # 随机生成一个 RGB 图像帧
+                    frame = np.random.randint(188, 256, (height, width, channels), dtype=np.uint8)
+                    
+                    # 将帧放入队列
+                    frame_queue.put(frame)
+                    # print("生成一帧数据，形状:", frame.shape, frame_queue.qsize())
+                    
+                    # 模拟实时性：避免过度消耗 CPU
+                    time.sleep(0.03)  # 每秒约生成 30 帧
+                except Exception as e:
+                    print(f"生成帧时出错: {e}")
+                    break
+        thread = threading.Thread(target=generate_frames, daemon=True)
+        thread.start()
 
 if __name__ == "__main__":
     demo.launch()
+
+
+

demo/echo_audio/README.md

@@ -0,0 +1,15 @@
+---
+title: Echo Audio
+emoji: 🪩
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Simple echo stream - simplest FastRTC demo
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/echo_audio/app.py

@@ -0,0 +1,45 @@
+import numpy as np
+from fastapi import FastAPI
+from fastapi.responses import RedirectResponse
+from fastrtc import ReplyOnPause, Stream, get_twilio_turn_credentials
+from gradio.utils import get_space
+
+
+def detection(audio: tuple[int, np.ndarray]):
+    # Implement any iterator that yields audio
+    # See "LLM Voice Chat" for a more complete example
+    yield audio
+
+
+stream = Stream(
+    handler=ReplyOnPause(detection),
+    modality="audio",
+    mode="send-receive",
+    rtc_configuration=get_twilio_turn_credentials() if get_space() else None,
+    concurrency_limit=5 if get_space() else None,
+    time_limit=90 if get_space() else None,
+)
+
+app = FastAPI()
+
+stream.mount(app)
+
+
+@app.get("/")
+async def index():
+    return RedirectResponse(
+        url="/ui" if not get_space() else "https://fastrtc-echo-audio.hf.space/ui/"
+    )
+
+
+if __name__ == "__main__":
+    import os
+
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        stream.fastphone(port=7860)
+    else:
+        import uvicorn
+
+        uvicorn.run(app, host="0.0.0.0", port=7860)

demo/echo_audio/requirements.txt

@@ -0,0 +1,3 @@
+fastrtc[vad]
+twilio
+python-dotenv

demo/gemini_audio_video/README.md

@@ -0,0 +1,15 @@
+---
+title: Gemini Audio Video
+emoji: ♊️
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Gemini understands audio and video!
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GEMINI_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/gemini_audio_video/app.py

@@ -0,0 +1,185 @@
+import asyncio
+import base64
+import os
+import time
+from io import BytesIO
+
+import gradio as gr
+import numpy as np
+from dotenv import load_dotenv
+from fastrtc import (
+    AsyncAudioVideoStreamHandler,
+    Stream,
+    WebRTC,
+    get_twilio_turn_credentials,
+)
+from google import genai
+from gradio.utils import get_space
+from PIL import Image
+
+load_dotenv()
+
+
+def encode_audio(data: np.ndarray) -> dict:
+    """Encode Audio data to send to the server"""
+    return {
+        "mime_type": "audio/pcm",
+        "data": base64.b64encode(data.tobytes()).decode("UTF-8"),
+    }
+
+
+def encode_image(data: np.ndarray) -> dict:
+    with BytesIO() as output_bytes:
+        pil_image = Image.fromarray(data)
+        pil_image.save(output_bytes, "JPEG")
+        bytes_data = output_bytes.getvalue()
+    base64_str = str(base64.b64encode(bytes_data), "utf-8")
+    return {"mime_type": "image/jpeg", "data": base64_str}
+
+
+class GeminiHandler(AsyncAudioVideoStreamHandler):
+    def __init__(
+        self,
+    ) -> None:
+        super().__init__(
+            "mono",
+            output_sample_rate=24000,
+            output_frame_size=480,
+            input_sample_rate=16000,
+        )
+        self.audio_queue = asyncio.Queue()
+        self.video_queue = asyncio.Queue()
+        self.quit = asyncio.Event()
+        self.session = None
+        self.last_frame_time = 0
+        self.quit = asyncio.Event()
+
+    def copy(self) -> "GeminiHandler":
+        return GeminiHandler()
+
+    async def start_up(self):
+        client = genai.Client(
+            api_key=os.getenv("GEMINI_API_KEY"), http_options={"api_version": "v1alpha"}
+        )
+        config = {"response_modalities": ["AUDIO"]}
+        async with client.aio.live.connect(
+            model="gemini-2.0-flash-exp", config=config
+        ) as session:
+            self.session = session
+            print("set session")
+            while not self.quit.is_set():
+                turn = self.session.receive()
+                async for response in turn:
+                    if data := response.data:
+                        audio = np.frombuffer(data, dtype=np.int16).reshape(1, -1)
+                        self.audio_queue.put_nowait(audio)
+
+    async def video_receive(self, frame: np.ndarray):
+        if self.session:
+            # send image every 1 second
+            print(time.time() - self.last_frame_time)
+            if time.time() - self.last_frame_time > 1:
+                self.last_frame_time = time.time()
+                await self.session.send(input=encode_image(frame))
+                if self.latest_args[1] is not None:
+                    await self.session.send(input=encode_image(self.latest_args[1]))
+
+        self.video_queue.put_nowait(frame)
+
+    async def video_emit(self):
+        return await self.video_queue.get()
+
+    async def receive(self, frame: tuple[int, np.ndarray]) -> None:
+        _, array = frame
+        array = array.squeeze()
+        audio_message = encode_audio(array)
+        if self.session:
+            await self.session.send(input=audio_message)
+
+    async def emit(self):
+        array = await self.audio_queue.get()
+        return (self.output_sample_rate, array)
+
+    async def shutdown(self) -> None:
+        if self.session:
+            self.quit.set()
+            await self.session._websocket.close()
+            self.quit.clear()
+
+
+stream = Stream(
+    handler=GeminiHandler(),
+    modality="audio-video",
+    mode="send-receive",
+    rtc_configuration=get_twilio_turn_credentials()
+    if get_space() == "spaces"
+    else None,
+    time_limit=90 if get_space() else None,
+    additional_inputs=[
+        gr.Image(label="Image", type="numpy", sources=["upload", "clipboard"])
+    ],
+    ui_args={
+        "icon": "https://www.gstatic.com/lamda/images/gemini_favicon_f069958c85030456e93de685481c559f160ea06b.png",
+        "pulse_color": "rgb(255, 255, 255)",
+        "icon_button_color": "rgb(255, 255, 255)",
+        "title": "Gemini Audio Video Chat",
+    },
+)
+
+css = """
+#video-source {max-width: 600px !important; max-height: 600 !important;}
+"""
+
+with gr.Blocks(css=css) as demo:
+    gr.HTML(
+        """
+    <div style='display: flex; align-items: center; justify-content: center; gap: 20px'>
+        <div style="background-color: var(--block-background-fill); border-radius: 8px">
+            <img src="https://www.gstatic.com/lamda/images/gemini_favicon_f069958c85030456e93de685481c559f160ea06b.png" style="width: 100px; height: 100px;">
+        </div>
+        <div>
+            <h1>Gen AI SDK Voice Chat</h1>
+            <p>Speak with Gemini using real-time audio + video streaming</p>
+            <p>Powered by <a href="https://gradio.app/">Gradio</a> and <a href=https://freddyaboulton.github.io/gradio-webrtc/">WebRTC</a>⚡️</p>
+            <p>Get an API Key <a href="https://support.google.com/googleapi/answer/6158862?hl=en">here</a></p>
+        </div>
+    </div>
+    """
+    )
+    with gr.Row() as row:
+        with gr.Column():
+            webrtc = WebRTC(
+                label="Video Chat",
+                modality="audio-video",
+                mode="send-receive",
+                elem_id="video-source",
+                rtc_configuration=get_twilio_turn_credentials()
+                if get_space() == "spaces"
+                else None,
+                icon="https://www.gstatic.com/lamda/images/gemini_favicon_f069958c85030456e93de685481c559f160ea06b.png",
+                pulse_color="rgb(255, 255, 255)",
+                icon_button_color="rgb(255, 255, 255)",
+            )
+        with gr.Column():
+            image_input = gr.Image(
+                label="Image", type="numpy", sources=["upload", "clipboard"]
+            )
+
+        webrtc.stream(
+            GeminiHandler(),
+            inputs=[webrtc, image_input],
+            outputs=[webrtc],
+            time_limit=60 if get_space() else None,
+            concurrency_limit=2 if get_space() else None,
+        )
+
+stream.ui = demo
+
+
+if __name__ == "__main__":
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        raise ValueError("Phone mode not supported for this demo")
+    else:
+        stream.ui.launch(server_port=7860)

demo/gemini_audio_video/requirements.txt

@@ -0,0 +1,4 @@
+fastrtc
+python-dotenv
+google-genai
+twilio

demo/gemini_conversation/README.md

@@ -0,0 +1,15 @@
+---
+title: Gemini Talking to Gemini
+emoji: ♊️
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.17.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Have two Gemini agents talk to each other
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GEMINI_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/gemini_conversation/app.py

@@ -0,0 +1,232 @@
+import asyncio
+import base64
+import os
+from pathlib import Path
+from typing import AsyncGenerator
+
+import librosa
+import numpy as np
+from dotenv import load_dotenv
+from fastrtc import (
+    AsyncStreamHandler,
+    Stream,
+    get_tts_model,
+    wait_for_item,
+)
+from fastrtc.utils import audio_to_int16
+from google import genai
+from google.genai.types import (
+    Content,
+    LiveConnectConfig,
+    Part,
+    PrebuiltVoiceConfig,
+    SpeechConfig,
+    VoiceConfig,
+)
+
+load_dotenv()
+
+cur_dir = Path(__file__).parent
+
+SAMPLE_RATE = 24000
+
+tts_model = get_tts_model()
+
+
+class GeminiHandler(AsyncStreamHandler):
+    """Handler for the Gemini API"""
+
+    def __init__(
+        self,
+    ) -> None:
+        super().__init__(
+            expected_layout="mono",
+            output_sample_rate=24000,
+            output_frame_size=480,
+            input_sample_rate=24000,
+        )
+        self.input_queue: asyncio.Queue = asyncio.Queue()
+        self.output_queue: asyncio.Queue = asyncio.Queue()
+        self.quit: asyncio.Event = asyncio.Event()
+
+    def copy(self) -> "GeminiHandler":
+        return GeminiHandler()
+
+    async def start_up(self):
+        voice_name = "Charon"
+        client = genai.Client(
+            api_key=os.getenv("GEMINI_API_KEY"),
+            http_options={"api_version": "v1alpha"},
+        )
+
+        config = LiveConnectConfig(
+            response_modalities=["AUDIO"],  # type: ignore
+            speech_config=SpeechConfig(
+                voice_config=VoiceConfig(
+                    prebuilt_voice_config=PrebuiltVoiceConfig(
+                        voice_name=voice_name,
+                    )
+                )
+            ),
+            system_instruction=Content(
+                parts=[Part(text="You are a helpful assistant.")],
+                role="system",
+            ),
+        )
+        async with client.aio.live.connect(
+            model="gemini-2.0-flash-exp", config=config
+        ) as session:
+            async for audio in session.start_stream(
+                stream=self.stream(), mime_type="audio/pcm"
+            ):
+                if audio.data:
+                    array = np.frombuffer(audio.data, dtype=np.int16)
+                    self.output_queue.put_nowait((self.output_sample_rate, array))
+
+    async def stream(self) -> AsyncGenerator[bytes, None]:
+        while not self.quit.is_set():
+            try:
+                audio = await asyncio.wait_for(self.input_queue.get(), 0.1)
+                yield audio
+            except (asyncio.TimeoutError, TimeoutError):
+                pass
+
+    async def receive(self, frame: tuple[int, np.ndarray]) -> None:
+        _, array = frame
+        array = array.squeeze()
+        audio_message = base64.b64encode(array.tobytes()).decode("UTF-8")
+        self.input_queue.put_nowait(audio_message)
+
+    async def emit(self) -> tuple[int, np.ndarray] | None:
+        return await wait_for_item(self.output_queue)
+
+    def shutdown(self) -> None:
+        self.quit.set()
+
+
+class GeminiHandler2(GeminiHandler):
+    async def start_up(self):
+        starting_message = tts_model.tts("Can you help me make an omelette?")
+        starting_message = librosa.resample(
+            starting_message[1],
+            orig_sr=starting_message[0],
+            target_sr=self.output_sample_rate,
+        )
+        starting_message = audio_to_int16((self.output_sample_rate, starting_message))
+        await self.output_queue.put((self.output_sample_rate, starting_message))
+        voice_name = "Puck"
+        client = genai.Client(
+            api_key=os.getenv("GEMINI_API_KEY"),
+            http_options={"api_version": "v1alpha"},
+        )
+
+        config = LiveConnectConfig(
+            response_modalities=["AUDIO"],  # type: ignore
+            speech_config=SpeechConfig(
+                voice_config=VoiceConfig(
+                    prebuilt_voice_config=PrebuiltVoiceConfig(
+                        voice_name=voice_name,
+                    )
+                )
+            ),
+            system_instruction=Content(
+                parts=[
+                    Part(
+                        text="You are a cooking student who wants to learn how to make an omelette."
+                    ),
+                    Part(
+                        text="You are currently in the kitchen with a teacher who is helping you make an omelette."
+                    ),
+                    Part(
+                        text="Please wait for the teacher to tell you what to do next. Follow the teacher's instructions carefully."
+                    ),
+                ],
+                role="system",
+            ),
+        )
+        async with client.aio.live.connect(
+            model="gemini-2.0-flash-exp", config=config
+        ) as session:
+            async for audio in session.start_stream(
+                stream=self.stream(), mime_type="audio/pcm"
+            ):
+                if audio.data:
+                    array = np.frombuffer(audio.data, dtype=np.int16)
+                    self.output_queue.put_nowait((self.output_sample_rate, array))
+
+    def copy(self) -> "GeminiHandler2":
+        return GeminiHandler2()
+
+
+gemini_stream = Stream(
+    GeminiHandler(),
+    modality="audio",
+    mode="send-receive",
+    ui_args={
+        "title": "Gemini Teacher",
+        "icon": "https://www.gstatic.com/lamda/images/gemini_favicon_f069958c85030456e93de685481c559f160ea06b.png",
+        "pulse_color": "rgb(74, 138, 213)",
+        "icon_button_color": "rgb(255, 255, 255)",
+    },
+)
+
+gemini_stream_2 = Stream(
+    GeminiHandler2(),
+    modality="audio",
+    mode="send-receive",
+    ui_args={
+        "title": "Gemini Student",
+        "icon": "https://www.gstatic.com/lamda/images/gemini_favicon_f069958c85030456e93de685481c559f160ea06b.png",
+        "pulse_color": "rgb(132, 112, 196)",
+        "icon_button_color": "rgb(255, 255, 255)",
+    },
+)
+
+if __name__ == "__main__":
+    import gradio as gr
+    from gradio.utils import get_space
+
+    if not get_space():
+        with gr.Blocks() as demo:
+            gr.HTML(
+                """
+                <div style="display: flex; justify-content: center; align-items: center;">
+                <h1>Gemini Conversation</h1>                
+                </div>
+                """
+            )
+            gr.Markdown(
+                """# How to run this demo
+                
+                - Clone the repo - top right of the page click the vertical three dots and select "Clone repository"
+                - Open the repo in a terminal and install the dependencies
+                - Get a gemini API key [here](https://ai.google.dev/gemini-api/docs/api-key)
+                - Create a `.env` file in the root of the repo and add the following:
+                ```
+                GEMINI_API_KEY=<your_gemini_api_key>
+                ```
+                - Run the app with `python app.py`
+                - This will print the two URLs of the agents running locally
+                - Use ngrok to exponse one agent to the internet. This is so that you can acces it from your phone
+                - Use the ngrok URL to access the agent from your phone
+                - Now, start the "teacher gemini" agent first. Then, start the "student gemini" agent. The student gemini will start talking to the teacher gemini. And the teacher gemini will respond!
+
+                Important:
+                - Make sure the audio sources are not too close to each other or too loud. Sometimes that causes them to talk over each other..
+                - Feel free to modify the `system_instruction` to change the behavior of the agents.
+                - You can also modify the `voice_name` to change the voice of the agents.
+                - Have fun!
+                """
+            )
+        demo.launch()
+
+    import time
+
+    _ = gemini_stream.ui.launch(server_port=7860, prevent_thread_lock=True)
+    _ = gemini_stream_2.ui.launch(server_port=7861, prevent_thread_lock=True)
+    try:
+        while True:
+            time.sleep(1)
+    except KeyboardInterrupt:
+        gemini_stream.ui.close()
+        gemini_stream_2.ui.close()

demo/hello_computer/README.md

@@ -0,0 +1,15 @@
+---
+title: Hello Computer
+emoji: 💻
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Say computer before asking your question
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|SAMBANOVA_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/hello_computer/README_gradio.md

@@ -0,0 +1,15 @@
+---
+title: Hello Computer (Gradio)
+emoji: 💻
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Say computer (Gradio)
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|SAMBANOVA_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/hello_computer/app.py

@@ -0,0 +1,145 @@
+import base64
+import json
+import os
+from pathlib import Path
+
+import gradio as gr
+import huggingface_hub
+import numpy as np
+from dotenv import load_dotenv
+from fastapi import FastAPI
+from fastapi.responses import HTMLResponse, StreamingResponse
+from fastrtc import (
+    AdditionalOutputs,
+    ReplyOnStopWords,
+    Stream,
+    get_stt_model,
+    get_twilio_turn_credentials,
+)
+from gradio.utils import get_space
+from pydantic import BaseModel
+
+load_dotenv()
+
+curr_dir = Path(__file__).parent
+
+
+client = huggingface_hub.InferenceClient(
+    api_key=os.environ.get("SAMBANOVA_API_KEY"),
+    provider="sambanova",
+)
+model = get_stt_model()
+
+
+def response(
+    audio: tuple[int, np.ndarray],
+    gradio_chatbot: list[dict] | None = None,
+    conversation_state: list[dict] | None = None,
+):
+    gradio_chatbot = gradio_chatbot or []
+    conversation_state = conversation_state or []
+    text = model.stt(audio)
+    print("STT in handler", text)
+    sample_rate, array = audio
+    gradio_chatbot.append(
+        {"role": "user", "content": gr.Audio((sample_rate, array.squeeze()))}
+    )
+    yield AdditionalOutputs(gradio_chatbot, conversation_state)
+
+    conversation_state.append({"role": "user", "content": text})
+
+    request = client.chat.completions.create(
+        model="meta-llama/Llama-3.2-3B-Instruct",
+        messages=conversation_state,  # type: ignore
+        temperature=0.1,
+        top_p=0.1,
+    )
+    response = {"role": "assistant", "content": request.choices[0].message.content}
+
+    conversation_state.append(response)
+    gradio_chatbot.append(response)
+
+    yield AdditionalOutputs(gradio_chatbot, conversation_state)
+
+
+chatbot = gr.Chatbot(type="messages", value=[])
+state = gr.State(value=[])
+stream = Stream(
+    ReplyOnStopWords(
+        response,  # type: ignore
+        stop_words=["computer"],
+        input_sample_rate=16000,
+    ),
+    mode="send",
+    modality="audio",
+    additional_inputs=[chatbot, state],
+    additional_outputs=[chatbot, state],
+    additional_outputs_handler=lambda *a: (a[2], a[3]),
+    concurrency_limit=5 if get_space() else None,
+    time_limit=90 if get_space() else None,
+    rtc_configuration=get_twilio_turn_credentials() if get_space() else None,
+)
+app = FastAPI()
+stream.mount(app)
+
+
+class Message(BaseModel):
+    role: str
+    content: str
+
+
+class InputData(BaseModel):
+    webrtc_id: str
+    chatbot: list[Message]
+    state: list[Message]
+
+
+@app.get("/")
+async def _():
+    rtc_config = get_twilio_turn_credentials() if get_space() else None
+    html_content = (curr_dir / "index.html").read_text()
+    html_content = html_content.replace("__RTC_CONFIGURATION__", json.dumps(rtc_config))
+    return HTMLResponse(content=html_content)
+
+
+@app.post("/input_hook")
+async def _(data: InputData):
+    body = data.model_dump()
+    stream.set_input(data.webrtc_id, body["chatbot"], body["state"])
+
+
+def audio_to_base64(file_path):
+    audio_format = "wav"
+    with open(file_path, "rb") as audio_file:
+        encoded_audio = base64.b64encode(audio_file.read()).decode("utf-8")
+    return f"data:audio/{audio_format};base64,{encoded_audio}"
+
+
+@app.get("/outputs")
+async def _(webrtc_id: str):
+    async def output_stream():
+        async for output in stream.output_stream(webrtc_id):
+            chatbot = output.args[0]
+            state = output.args[1]
+            data = {
+                "message": state[-1],
+                "audio": audio_to_base64(chatbot[-1]["content"].value["path"])
+                if chatbot[-1]["role"] == "user"
+                else None,
+            }
+            yield f"event: output\ndata: {json.dumps(data)}\n\n"
+
+    return StreamingResponse(output_stream(), media_type="text/event-stream")
+
+
+if __name__ == "__main__":
+    import os
+
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        raise ValueError("Phone mode not supported")
+    else:
+        import uvicorn
+
+        uvicorn.run(app, host="0.0.0.0", port=7860)

demo/hello_computer/index.html

@@ -0,0 +1,486 @@
+<!DOCTYPE html>
+<html lang="en">
+
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Hello Computer 💻</title>
+    <style>
+        body {
+            font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;
+            background-color: #f8f9fa;
+            color: #1a1a1a;
+            margin: 0;
+            padding: 20px;
+            height: 100vh;
+            box-sizing: border-box;
+        }
+
+        .container {
+            max-width: 800px;
+            margin: 0 auto;
+            height: calc(100% - 100px);
+        }
+
+        .logo {
+            text-align: center;
+            margin-bottom: 40px;
+        }
+
+        .chat-container {
+            background: white;
+            border-radius: 8px;
+            box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1);
+            padding: 20px;
+            height: 90%;
+            box-sizing: border-box;
+            display: flex;
+            flex-direction: column;
+        }
+
+        .chat-messages {
+            flex-grow: 1;
+            overflow-y: auto;
+            margin-bottom: 20px;
+            padding: 10px;
+        }
+
+        .message {
+            margin-bottom: 20px;
+            padding: 12px;
+            border-radius: 8px;
+            font-size: 14px;
+            line-height: 1.5;
+        }
+
+        .message.user {
+            background-color: #e9ecef;
+            margin-left: 20%;
+        }
+
+        .message.assistant {
+            background-color: #f1f3f5;
+            margin-right: 20%;
+        }
+
+        .controls {
+            text-align: center;
+            margin-top: 20px;
+        }
+
+        button {
+            background-color: #0066cc;
+            color: white;
+            border: none;
+            padding: 12px 24px;
+            font-family: inherit;
+            font-size: 14px;
+            cursor: pointer;
+            transition: all 0.3s;
+            border-radius: 4px;
+            font-weight: 500;
+        }
+
+        button:hover {
+            background-color: #0052a3;
+        }
+
+        #audio-output {
+            display: none;
+        }
+
+        .icon-with-spinner {
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            gap: 12px;
+            min-width: 180px;
+        }
+
+        .spinner {
+            width: 20px;
+            height: 20px;
+            border: 2px solid #ffffff;
+            border-top-color: transparent;
+            border-radius: 50%;
+            animation: spin 1s linear infinite;
+            flex-shrink: 0;
+        }
+
+        @keyframes spin {
+            to {
+                transform: rotate(360deg);
+            }
+        }
+
+        .pulse-container {
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            gap: 12px;
+            min-width: 180px;
+        }
+
+        .pulse-circle {
+            width: 20px;
+            height: 20px;
+            border-radius: 50%;
+            background-color: #ffffff;
+            opacity: 0.2;
+            flex-shrink: 0;
+            transform: translateX(-0%) scale(var(--audio-level, 1));
+            transition: transform 0.1s ease;
+        }
+
+        /* Add styles for typing indicator */
+        .typing-indicator {
+            padding: 8px;
+            background-color: #f1f3f5;
+            border-radius: 8px;
+            margin-bottom: 10px;
+            display: none;
+        }
+
+        .dots {
+            display: inline-flex;
+            gap: 4px;
+        }
+
+        .dot {
+            width: 8px;
+            height: 8px;
+            background-color: #0066cc;
+            border-radius: 50%;
+            animation: pulse 1.5s infinite;
+            opacity: 0.5;
+        }
+
+        .dot:nth-child(2) {
+            animation-delay: 0.5s;
+        }
+
+        .dot:nth-child(3) {
+            animation-delay: 1s;
+        }
+
+        @keyframes pulse {
+
+            0%,
+            100% {
+                opacity: 0.5;
+                transform: scale(1);
+            }
+
+            50% {
+                opacity: 1;
+                transform: scale(1.2);
+            }
+        }
+
+        /* Add styles for toast notifications */
+        .toast {
+            position: fixed;
+            top: 20px;
+            left: 50%;
+            transform: translateX(-50%);
+            padding: 16px 24px;
+            border-radius: 4px;
+            font-size: 14px;
+            z-index: 1000;
+            display: none;
+            box-shadow: 0 2px 5px rgba(0, 0, 0, 0.2);
+        }
+
+        .toast.error {
+            background-color: #f44336;
+            color: white;
+        }
+
+        .toast.warning {
+            background-color: #ffd700;
+            color: black;
+        }
+    </style>
+</head>
+
+<body>
+    <!-- Add toast element after body opening tag -->
+    <div id="error-toast" class="toast"></div>
+    <div class="container">
+        <div class="logo">
+            <h1>Hello Computer 💻</h1>
+            <h2 style="font-size: 1.2em; color: #666; margin-top: 10px;">Say 'Computer' before asking your question</h2>
+        </div>
+        <div class="chat-container">
+            <div class="chat-messages" id="chat-messages"></div>
+            <div class="typing-indicator" id="typing-indicator">
+                <div class="dots">
+                    <div class="dot"></div>
+                    <div class="dot"></div>
+                    <div class="dot"></div>
+                </div>
+            </div>
+        </div>
+        <div class="controls">
+            <button id="start-button">Start Conversation</button>
+        </div>
+    </div>
+    <audio id="audio-output"></audio>
+
+    <script>
+        let peerConnection;
+        let webrtc_id;
+        const startButton = document.getElementById('start-button');
+        const chatMessages = document.getElementById('chat-messages');
+
+        let audioLevel = 0;
+        let animationFrame;
+        let audioContext, analyser, audioSource;
+        let messages = [];
+        let eventSource;
+
+        function updateButtonState() {
+            const button = document.getElementById('start-button');
+            if (peerConnection && (peerConnection.connectionState === 'connecting' || peerConnection.connectionState === 'new')) {
+                button.innerHTML = `
+                    <div class="icon-with-spinner">
+                        <div class="spinner"></div>
+                        <span>Connecting...</span>
+                    </div>
+                `;
+            } else if (peerConnection && peerConnection.connectionState === 'connected') {
+                button.innerHTML = `
+                    <div class="pulse-container">
+                        <div class="pulse-circle"></div>
+                        <span>Stop Conversation</span>
+                    </div>
+                `;
+            } else {
+                button.innerHTML = 'Start Conversation';
+            }
+        }
+
+        function setupAudioVisualization(stream) {
+            audioContext = new (window.AudioContext || window.webkitAudioContext)();
+            analyser = audioContext.createAnalyser();
+            audioSource = audioContext.createMediaStreamSource(stream);
+            audioSource.connect(analyser);
+            analyser.fftSize = 64;
+            const dataArray = new Uint8Array(analyser.frequencyBinCount);
+
+            function updateAudioLevel() {
+                analyser.getByteFrequencyData(dataArray);
+                const average = Array.from(dataArray).reduce((a, b) => a + b, 0) / dataArray.length;
+                audioLevel = average / 255;
+
+                const pulseCircle = document.querySelector('.pulse-circle');
+                if (pulseCircle) {
+                    pulseCircle.style.setProperty('--audio-level', 1 + audioLevel);
+                }
+
+                animationFrame = requestAnimationFrame(updateAudioLevel);
+            }
+            updateAudioLevel();
+        }
+
+        function showError(message) {
+            const toast = document.getElementById('error-toast');
+            toast.textContent = message;
+            toast.className = 'toast error';
+            toast.style.display = 'block';
+
+            // Hide toast after 5 seconds
+            setTimeout(() => {
+                toast.style.display = 'none';
+            }, 5000);
+        }
+
+        function handleMessage(event) {
+            const eventJson = JSON.parse(event.data);
+            const typingIndicator = document.getElementById('typing-indicator');
+
+            if (eventJson.type === "error") {
+                showError(eventJson.message);
+            } else if (eventJson.type === "send_input") {
+                fetch('/input_hook', {
+                    method: 'POST',
+                    headers: {
+                        'Content-Type': 'application/json',
+                    },
+                    body: JSON.stringify({
+                        webrtc_id: webrtc_id,
+                        chatbot: messages,
+                        state: messages
+                    })
+                });
+            } else if (eventJson.type === "log") {
+                if (eventJson.data === "pause_detected") {
+                    typingIndicator.style.display = 'block';
+                    chatMessages.scrollTop = chatMessages.scrollHeight;
+                } else if (eventJson.data === "response_starting") {
+                    typingIndicator.style.display = 'none';
+                }
+            }
+        }
+
+        async function setupWebRTC() {
+            const config = __RTC_CONFIGURATION__;
+            peerConnection = new RTCPeerConnection(config);
+
+            const timeoutId = setTimeout(() => {
+                const toast = document.getElementById('error-toast');
+                toast.textContent = "Connection is taking longer than usual. Are you on a VPN?";
+                toast.className = 'toast warning';
+                toast.style.display = 'block';
+
+                // Hide warning after 5 seconds
+                setTimeout(() => {
+                    toast.style.display = 'none';
+                }, 5000);
+            }, 5000);
+
+            try {
+                const stream = await navigator.mediaDevices.getUserMedia({
+                    audio: true
+                });
+
+                setupAudioVisualization(stream);
+
+                stream.getTracks().forEach(track => {
+                    peerConnection.addTrack(track, stream);
+                });
+
+                const dataChannel = peerConnection.createDataChannel('text');
+                dataChannel.onmessage = handleMessage;
+
+                const offer = await peerConnection.createOffer();
+                await peerConnection.setLocalDescription(offer);
+
+                await new Promise((resolve) => {
+                    if (peerConnection.iceGatheringState === "complete") {
+                        resolve();
+                    } else {
+                        const checkState = () => {
+                            if (peerConnection.iceGatheringState === "complete") {
+                                peerConnection.removeEventListener("icegatheringstatechange", checkState);
+                                resolve();
+                            }
+                        };
+                        peerConnection.addEventListener("icegatheringstatechange", checkState);
+                    }
+                });
+
+                peerConnection.addEventListener('connectionstatechange', () => {
+                    console.log('connectionstatechange', peerConnection.connectionState);
+                    if (peerConnection.connectionState === 'connected') {
+                        clearTimeout(timeoutId);
+                        const toast = document.getElementById('error-toast');
+                        toast.style.display = 'none';
+                    }
+                    updateButtonState();
+                });
+
+                webrtc_id = Math.random().toString(36).substring(7);
+
+                const response = await fetch('/webrtc/offer', {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify({
+                        sdp: peerConnection.localDescription.sdp,
+                        type: peerConnection.localDescription.type,
+                        webrtc_id: webrtc_id
+                    })
+                });
+
+                const serverResponse = await response.json();
+
+                if (serverResponse.status === 'failed') {
+                    showError(serverResponse.meta.error === 'concurrency_limit_reached'
+                        ? `Too many connections. Maximum limit is ${serverResponse.meta.limit}`
+                        : serverResponse.meta.error);
+                    stop();
+                    return;
+                }
+
+                await peerConnection.setRemoteDescription(serverResponse);
+
+                eventSource = new EventSource('/outputs?webrtc_id=' + webrtc_id);
+                eventSource.addEventListener("output", (event) => {
+                    const eventJson = JSON.parse(event.data);
+                    console.log(eventJson);
+                    messages.push(eventJson.message);
+                    addMessage(eventJson.message.role, eventJson.audio ?? eventJson.message.content);
+                });
+            } catch (err) {
+                clearTimeout(timeoutId);
+                console.error('Error setting up WebRTC:', err);
+                showError('Failed to establish connection. Please try again.');
+                stop();
+            }
+        }
+
+        function addMessage(role, content) {
+            const messageDiv = document.createElement('div');
+            messageDiv.classList.add('message', role);
+
+            if (role === 'user') {
+                // Create audio element for user messages
+                const audio = document.createElement('audio');
+                audio.controls = true;
+                audio.src = content;
+                messageDiv.appendChild(audio);
+            } else {
+                // Text content for assistant messages
+                messageDiv.textContent = content;
+            }
+
+            chatMessages.appendChild(messageDiv);
+            chatMessages.scrollTop = chatMessages.scrollHeight;
+        }
+
+        function stop() {
+            if (eventSource) {
+                eventSource.close();
+                eventSource = null;
+            }
+
+            if (animationFrame) {
+                cancelAnimationFrame(animationFrame);
+            }
+            if (audioContext) {
+                audioContext.close();
+                audioContext = null;
+                analyser = null;
+                audioSource = null;
+            }
+            if (peerConnection) {
+                if (peerConnection.getTransceivers) {
+                    peerConnection.getTransceivers().forEach(transceiver => {
+                        if (transceiver.stop) {
+                            transceiver.stop();
+                        }
+                    });
+                }
+
+                if (peerConnection.getSenders) {
+                    peerConnection.getSenders().forEach(sender => {
+                        if (sender.track && sender.track.stop) sender.track.stop();
+                    });
+                }
+                peerConnection.close();
+            }
+            updateButtonState();
+            audioLevel = 0;
+        }
+
+        startButton.addEventListener('click', () => {
+            if (!peerConnection || peerConnection.connectionState !== 'connected') {
+                setupWebRTC();
+            } else {
+                stop();
+            }
+        });
+    </script>
+</body>
+
+</html>

demo/hello_computer/requirements.txt

@@ -0,0 +1,4 @@
+fastrtc[stopword]
+python-dotenv
+huggingface_hub>=0.29.0
+twilio

demo/llama_code_editor/README.md

@@ -0,0 +1,16 @@
+---
+title: Llama Code Editor
+emoji: 🦙
+colorFrom: indigo
+colorTo: pink
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Create interactive HTML web pages with your voice
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN,
+secret|SAMBANOVA_API_KEY, secret|GROQ_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/llama_code_editor/app.py

@@ -0,0 +1,45 @@
+from fastapi import FastAPI
+from fastapi.responses import RedirectResponse
+from fastrtc import Stream
+from gradio.utils import get_space
+
+try:
+    from demo.llama_code_editor.handler import (
+        CodeHandler,
+    )
+    from demo.llama_code_editor.ui import demo as ui
+except (ImportError, ModuleNotFoundError):
+    from handler import CodeHandler
+    from ui import demo as ui
+
+
+stream = Stream(
+    handler=CodeHandler,
+    modality="audio",
+    mode="send-receive",
+    concurrency_limit=10 if get_space() else None,
+    time_limit=90 if get_space() else None,
+)
+
+stream.ui = ui
+
+app = FastAPI()
+
+
+@app.get("/")
+async def _():
+    url = "/ui" if not get_space() else "https://fastrtc-llama-code-editor.hf.space/ui/"
+    return RedirectResponse(url)
+
+
+if __name__ == "__main__":
+    import os
+
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860, server_name="0.0.0.0")
+    elif mode == "PHONE":
+        stream.fastphone(host="0.0.0.0", port=7860)
+    else:
+        import uvicorn
+
+        uvicorn.run(app, host="0.0.0.0", port=7860)

demo/llama_code_editor/assets/sandbox.html

@@ -0,0 +1,37 @@
+<div style="
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  min-height: 400px;
+  background: linear-gradient(135deg, #f5f7fa 0%, #e4e8ec 100%);
+  border-radius: 8px;
+  border: 2px dashed #cbd5e1;
+  padding: 2rem;
+  text-align: center;
+  color: #64748b;
+  font-family: system-ui, -apple-system, sans-serif;
+">
+  <div style="
+    width: 80px;
+    height: 80px;
+    margin-bottom: 1.5rem;
+    border: 3px solid #cbd5e1;
+    border-radius: 12px;
+    position: relative;
+  ">
+    <div style="
+      position: absolute;
+      top: 50%;
+      left: 50%;
+      transform: translate(-50%, -50%);
+      font-size: 2rem;
+    ">📦</div>
+  </div>
+  <h2 style="
+    margin: 0 0 0.5rem 0;
+    font-size: 1.5rem;
+    font-weight: 600;
+    color: #475569;
+  ">No Application Created</h2>
+</div>

demo/llama_code_editor/assets/spinner.html

@@ -0,0 +1,60 @@
+<div style="
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  min-height: 400px;
+  background: linear-gradient(135deg, #f8fafc 0%, #f1f5f9 100%);
+  border-radius: 8px;
+  padding: 2rem;
+  text-align: center;
+  font-family: system-ui, -apple-system, sans-serif;
+">
+  <!-- Spinner container -->
+  <div style="
+    position: relative;
+    width: 64px;
+    height: 64px;
+    margin-bottom: 1.5rem;
+  ">
+    <!-- Static ring -->
+    <div style="
+      position: absolute;
+      width: 100%;
+      height: 100%;
+      border: 4px solid #e2e8f0;
+      border-radius: 50%;
+    "></div>
+    <!-- Animated spinner -->
+    <div style="
+      position: absolute;
+      width: 100%;
+      height: 100%;
+      border: 4px solid transparent;
+      border-top-color: #3b82f6;
+      border-radius: 50%;
+      animation: spin 1s linear infinite;
+    "></div>
+  </div>
+
+  <!-- Text content -->
+  <h2 style="
+    margin: 0 0 0.5rem 0;
+    font-size: 1.25rem;
+    font-weight: 600;
+    color: #475569;
+  ">Generating your application...</h2>
+  
+  <p style="
+    margin: 0;
+    font-size: 0.875rem;
+    color: #64748b;
+  ">This may take a few moments</p>
+
+  <style>
+    @keyframes spin {
+      0% { transform: rotate(0deg); }
+      100% { transform: rotate(360deg); }
+    }
+  </style>
+</div>

demo/llama_code_editor/handler.py

@@ -0,0 +1,73 @@
+import base64
+import os
+import re
+from pathlib import Path
+
+import numpy as np
+import openai
+from dotenv import load_dotenv
+from fastrtc import (
+    AdditionalOutputs,
+    ReplyOnPause,
+    audio_to_bytes,
+)
+from groq import Groq
+
+load_dotenv()
+
+groq_client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
+
+client = openai.OpenAI(
+    api_key=os.environ.get("SAMBANOVA_API_KEY"),
+    base_url="https://api.sambanova.ai/v1",
+)
+
+path = Path(__file__).parent / "assets"
+
+spinner_html = open(path / "spinner.html").read()
+
+
+system_prompt = "You are an AI coding assistant. Your task is to write single-file HTML applications based on a user's request. Only return the necessary code. Include all necessary imports and styles. You may also be asked to edit your original response."
+user_prompt = "Please write a single-file HTML application to fulfill the following request.\nThe message:{user_message}\nCurrent code you have written:{code}"
+
+
+def extract_html_content(text):
+    """
+    Extract content including HTML tags.
+    """
+    match = re.search(r"<!DOCTYPE html>.*?</html>", text, re.DOTALL)
+    return match.group(0) if match else None
+
+
+def display_in_sandbox(code):
+    encoded_html = base64.b64encode(code.encode("utf-8")).decode("utf-8")
+    data_uri = f"data:text/html;charset=utf-8;base64,{encoded_html}"
+    return f'<iframe src="{data_uri}" width="100%" height="600px"></iframe>'
+
+
+def generate(user_message: tuple[int, np.ndarray], history: list[dict], code: str):
+    yield AdditionalOutputs(history, spinner_html)
+
+    text = groq_client.audio.transcriptions.create(
+        file=("audio-file.mp3", audio_to_bytes(user_message)),
+        model="whisper-large-v3-turbo",
+        response_format="verbose_json",
+    ).text
+
+    user_msg_formatted = user_prompt.format(user_message=text, code=code)
+    history.append({"role": "user", "content": user_msg_formatted})
+
+    response = client.chat.completions.create(
+        model="Meta-Llama-3.1-70B-Instruct",
+        messages=history,  # type: ignore
+        temperature=0.1,
+        top_p=0.1,
+    )
+
+    output = response.choices[0].message.content
+    html_code = extract_html_content(output)
+    history.append({"role": "assistant", "content": output})
+    yield AdditionalOutputs(history, html_code)
+
+
+CodeHandler = ReplyOnPause(generate)  # type: ignore

demo/llama_code_editor/requirements.in

@@ -0,0 +1,5 @@
+fastrtc[vad]
+groq
+openai
+python-dotenv
+twilio

demo/llama_code_editor/requirements.txt

@@ -0,0 +1,295 @@
+# This file was autogenerated by uv via the following command:
+#    uv pip compile demo/llama_code_editor/requirements.in -o demo/llama_code_editor/requirements.txt
+aiofiles==23.2.1
+    # via gradio
+aiohappyeyeballs==2.4.6
+    # via aiohttp
+aiohttp==3.11.12
+    # via
+    #   aiohttp-retry
+    #   twilio
+aiohttp-retry==2.9.1
+    # via twilio
+aioice==0.9.0
+    # via aiortc
+aiortc==1.10.1
+    # via fastrtc
+aiosignal==1.3.2
+    # via aiohttp
+annotated-types==0.7.0
+    # via pydantic
+anyio==4.6.2.post1
+    # via
+    #   gradio
+    #   groq
+    #   httpx
+    #   openai
+    #   starlette
+attrs==25.1.0
+    # via aiohttp
+audioread==3.0.1
+    # via librosa
+av==12.3.0
+    # via aiortc
+certifi==2024.8.30
+    # via
+    #   httpcore
+    #   httpx
+    #   requests
+cffi==1.17.1
+    # via
+    #   aiortc
+    #   cryptography
+    #   pylibsrtp
+    #   soundfile
+charset-normalizer==3.4.0
+    # via requests
+click==8.1.7
+    # via
+    #   typer
+    #   uvicorn
+coloredlogs==15.0.1
+    # via onnxruntime
+cryptography==43.0.3
+    # via
+    #   aiortc
+    #   pyopenssl
+decorator==5.1.1
+    # via librosa
+distro==1.9.0
+    # via
+    #   groq
+    #   openai
+dnspython==2.7.0
+    # via aioice
+fastapi==0.115.5
+    # via gradio
+fastrtc==0.0.2.post4
+    # via -r demo/llama_code_editor/requirements.in
+ffmpy==0.4.0
+    # via gradio
+filelock==3.16.1
+    # via huggingface-hub
+flatbuffers==24.3.25
+    # via onnxruntime
+frozenlist==1.5.0
+    # via
+    #   aiohttp
+    #   aiosignal
+fsspec==2024.10.0
+    # via
+    #   gradio-client
+    #   huggingface-hub
+google-crc32c==1.6.0
+    # via aiortc
+gradio==5.16.0
+    # via fastrtc
+gradio-client==1.7.0
+    # via gradio
+groq==0.18.0
+    # via -r demo/llama_code_editor/requirements.in
+h11==0.14.0
+    # via
+    #   httpcore
+    #   uvicorn
+httpcore==1.0.7
+    # via httpx
+httpx==0.27.2
+    # via
+    #   gradio
+    #   gradio-client
+    #   groq
+    #   openai
+    #   safehttpx
+huggingface-hub==0.28.1
+    # via
+    #   gradio
+    #   gradio-client
+humanfriendly==10.0
+    # via coloredlogs
+idna==3.10
+    # via
+    #   anyio
+    #   httpx
+    #   requests
+    #   yarl
+ifaddr==0.2.0
+    # via aioice
+jinja2==3.1.4
+    # via gradio
+jiter==0.7.1
+    # via openai
+joblib==1.4.2
+    # via
+    #   librosa
+    #   scikit-learn
+lazy-loader==0.4
+    # via librosa
+librosa==0.10.2.post1
+    # via fastrtc
+llvmlite==0.43.0
+    # via numba
+markdown-it-py==3.0.0
+    # via rich
+markupsafe==2.1.5
+    # via
+    #   gradio
+    #   jinja2
+mdurl==0.1.2
+    # via markdown-it-py
+mpmath==1.3.0
+    # via sympy
+msgpack==1.1.0
+    # via librosa
+multidict==6.1.0
+    # via
+    #   aiohttp
+    #   yarl
+numba==0.60.0
+    # via librosa
+numpy==2.0.2
+    # via
+    #   gradio
+    #   librosa
+    #   numba
+    #   onnxruntime
+    #   pandas
+    #   scikit-learn
+    #   scipy
+    #   soxr
+onnxruntime==1.20.1
+    # via fastrtc
+openai==1.54.4
+    # via -r demo/llama_code_editor/requirements.in
+orjson==3.10.11
+    # via gradio
+packaging==24.2
+    # via
+    #   gradio
+    #   gradio-client
+    #   huggingface-hub
+    #   lazy-loader
+    #   onnxruntime
+    #   pooch
+pandas==2.2.3
+    # via gradio
+pillow==11.0.0
+    # via gradio
+platformdirs==4.3.6
+    # via pooch
+pooch==1.8.2
+    # via librosa
+propcache==0.2.1
+    # via
+    #   aiohttp
+    #   yarl
+protobuf==5.28.3
+    # via onnxruntime
+pycparser==2.22
+    # via cffi
+pydantic==2.9.2
+    # via
+    #   fastapi
+    #   gradio
+    #   groq
+    #   openai
+pydantic-core==2.23.4
+    # via pydantic
+pydub==0.25.1
+    # via gradio
+pyee==12.1.1
+    # via aiortc
+pygments==2.18.0
+    # via rich
+pyjwt==2.10.1
+    # via twilio
+pylibsrtp==0.10.0
+    # via aiortc
+pyopenssl==24.2.1
+    # via aiortc
+python-dateutil==2.9.0.post0
+    # via pandas
+python-dotenv==1.0.1
+    # via -r demo/llama_code_editor/requirements.in
+python-multipart==0.0.20
+    # via gradio
+pytz==2024.2
+    # via pandas
+pyyaml==6.0.2
+    # via
+    #   gradio
+    #   huggingface-hub
+requests==2.32.3
+    # via
+    #   huggingface-hub
+    #   pooch
+    #   twilio
+rich==13.9.4
+    # via typer
+ruff==0.9.6
+    # via gradio
+safehttpx==0.1.6
+    # via gradio
+scikit-learn==1.5.2
+    # via librosa
+scipy==1.14.1
+    # via
+    #   librosa
+    #   scikit-learn
+semantic-version==2.10.0
+    # via gradio
+shellingham==1.5.4
+    # via typer
+six==1.16.0
+    # via python-dateutil
+sniffio==1.3.1
+    # via
+    #   anyio
+    #   groq
+    #   httpx
+    #   openai
+soundfile==0.12.1
+    # via librosa
+soxr==0.5.0.post1
+    # via librosa
+starlette==0.42.0
+    # via
+    #   fastapi
+    #   gradio
+sympy==1.13.3
+    # via onnxruntime
+threadpoolctl==3.5.0
+    # via scikit-learn
+tomlkit==0.12.0
+    # via gradio
+tqdm==4.67.0
+    # via
+    #   huggingface-hub
+    #   openai
+twilio==9.4.5
+    # via -r demo/llama_code_editor/requirements.in
+typer==0.13.1
+    # via gradio
+typing-extensions==4.12.2
+    # via
+    #   fastapi
+    #   gradio
+    #   gradio-client
+    #   groq
+    #   huggingface-hub
+    #   librosa
+    #   openai
+    #   pydantic
+    #   pydantic-core
+    #   pyee
+    #   typer
+tzdata==2024.2
+    # via pandas
+urllib3==2.2.3
+    # via requests
+uvicorn==0.32.0
+    # via gradio
+websockets==12.0
+    # via gradio-client
+yarl==1.18.3
+    # via aiohttp

demo/llama_code_editor/ui.py

@@ -0,0 +1,75 @@
+from pathlib import Path
+
+import gradio as gr
+from dotenv import load_dotenv
+from fastrtc import WebRTC, get_twilio_turn_credentials
+from gradio.utils import get_space
+
+try:
+    from demo.llama_code_editor.handler import (
+        CodeHandler,
+        display_in_sandbox,
+        system_prompt,
+    )
+except (ImportError, ModuleNotFoundError):
+    from handler import CodeHandler, display_in_sandbox, system_prompt
+
+load_dotenv()
+
+path = Path(__file__).parent / "assets"
+
+with gr.Blocks(css=".code-component {max-height: 500px !important}") as demo:
+    history = gr.State([{"role": "system", "content": system_prompt}])
+    with gr.Row():
+        with gr.Column(scale=1):
+            gr.HTML(
+                """
+                <h1 style='text-align: center'>
+                Llama Code Editor
+                </h1>
+                <h2 style='text-align: center'>
+                Powered by SambaNova and Gradio-WebRTC ⚡️
+                </h2>
+                <p style='text-align: center'>
+                Create and edit single-file HTML applications with just your voice!
+                </p>
+                <p style='text-align: center'>
+                Each conversation is limited to 90 seconds. Once the time limit is up you can rejoin the conversation.
+                </p>
+                """
+            )
+            webrtc = WebRTC(
+                rtc_configuration=get_twilio_turn_credentials()
+                if get_space()
+                else None,
+                mode="send",
+                modality="audio",
+            )
+        with gr.Column(scale=10):
+            with gr.Tabs():
+                with gr.Tab("Sandbox"):
+                    sandbox = gr.HTML(value=open(path / "sandbox.html").read())
+                with gr.Tab("Code"):
+                    code = gr.Code(
+                        language="html",
+                        max_lines=50,
+                        interactive=False,
+                        elem_classes="code-component",
+                    )
+                with gr.Tab("Chat"):
+                    cb = gr.Chatbot(type="messages")
+
+    webrtc.stream(
+        CodeHandler,
+        inputs=[webrtc, history, code],
+        outputs=[webrtc],
+        time_limit=90 if get_space() else None,
+        concurrency_limit=10 if get_space() else None,
+    )
+    webrtc.on_additional_outputs(
+        lambda history, code: (history, code, history), outputs=[history, code, cb]
+    )
+    code.change(display_in_sandbox, code, sandbox, queue=False)
+
+if __name__ == "__main__":
+    demo.launch()

demo/llm_voice_chat/README.md

@@ -0,0 +1,15 @@
+---
+title: LLM Voice Chat
+emoji: 💻
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Talk to an LLM with ElevenLabs
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GROQ_API_KEY, secret|ELEVENLABS_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/llm_voice_chat/README_gradio.md

@@ -0,0 +1,15 @@
+---
+title: LLM Voice Chat (Gradio)
+emoji: 💻
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: LLM Voice by ElevenLabs (Gradio)
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GROQ_API_KEY, secret|ELEVENLABS_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/llm_voice_chat/app.py

@@ -0,0 +1,97 @@
+import os
+import time
+
+import gradio as gr
+import numpy as np
+from dotenv import load_dotenv
+from elevenlabs import ElevenLabs
+from fastapi import FastAPI
+from fastrtc import (
+    AdditionalOutputs,
+    ReplyOnPause,
+    Stream,
+    get_stt_model,
+    get_twilio_turn_credentials,
+)
+from gradio.utils import get_space
+from groq import Groq
+from numpy.typing import NDArray
+
+load_dotenv()
+groq_client = Groq()
+tts_client = ElevenLabs(api_key=os.getenv("ELEVENLABS_API_KEY"))
+stt_model = get_stt_model()
+
+
+# See "Talk to Claude" in Cookbook for an example of how to keep
+# track of the chat history.
+def response(
+    audio: tuple[int, NDArray[np.int16 | np.float32]],
+    chatbot: list[dict] | None = None,
+):
+    chatbot = chatbot or []
+    messages = [{"role": d["role"], "content": d["content"]} for d in chatbot]
+    start = time.time()
+    text = stt_model.stt(audio)
+    print("transcription", time.time() - start)
+    print("prompt", text)
+    chatbot.append({"role": "user", "content": text})
+    yield AdditionalOutputs(chatbot)
+    messages.append({"role": "user", "content": text})
+    response_text = (
+        groq_client.chat.completions.create(
+            model="llama-3.1-8b-instant",
+            max_tokens=200,
+            messages=messages,  # type: ignore
+        )
+        .choices[0]
+        .message.content
+    )
+
+    chatbot.append({"role": "assistant", "content": response_text})
+
+    for i, chunk in enumerate(
+        tts_client.text_to_speech.convert_as_stream(
+            text=response_text,  # type: ignore
+            voice_id="JBFqnCBsd6RMkjVDRZzb",
+            model_id="eleven_multilingual_v2",
+            output_format="pcm_24000",
+        )
+    ):
+        if i == 0:
+            yield AdditionalOutputs(chatbot)
+        audio_array = np.frombuffer(chunk, dtype=np.int16).reshape(1, -1)
+        yield (24000, audio_array)
+
+
+chatbot = gr.Chatbot(type="messages")
+stream = Stream(
+    modality="audio",
+    mode="send-receive",
+    handler=ReplyOnPause(response, input_sample_rate=16000),
+    additional_outputs_handler=lambda a, b: b,
+    additional_inputs=[chatbot],
+    additional_outputs=[chatbot],
+    rtc_configuration=get_twilio_turn_credentials() if get_space() else None,
+    concurrency_limit=5 if get_space() else None,
+    time_limit=90 if get_space() else None,
+    ui_args={"title": "LLM Voice Chat (Powered by Groq, ElevenLabs, and WebRTC ⚡️)"},
+)
+
+# Mount the STREAM UI to the FastAPI app
+# Because I don't want to build the UI manually
+app = FastAPI()
+app = gr.mount_gradio_app(app, stream.ui, path="/")
+
+
+if __name__ == "__main__":
+    import os
+
+    os.environ["GRADIO_SSR_MODE"] = "false"
+
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        stream.fastphone(host="0.0.0.0", port=7860)
+    else:
+        stream.ui.launch(server_port=7860)

demo/llm_voice_chat/requirements.txt

@@ -0,0 +1,6 @@
+fastrtc[stopword]
+python-dotenv
+openai
+twilio
+groq
+elevenlabs

demo/moonshine_live/README.md

@@ -0,0 +1,16 @@
+---
+title: Moonshine Live Transcription
+emoji: 🌕
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.17.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Real-time captions with Moonshine ONNX
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN]
+models: [onnx-community/moonshine-base-ONNX, UsefulSensors/moonshine-base]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/moonshine_live/app.py

@@ -0,0 +1,73 @@
+from functools import lru_cache
+from typing import Generator, Literal
+
+import gradio as gr
+import numpy as np
+from dotenv import load_dotenv
+from fastrtc import (
+    AdditionalOutputs,
+    ReplyOnPause,
+    Stream,
+    audio_to_float32,
+    get_twilio_turn_credentials,
+)
+from moonshine_onnx import MoonshineOnnxModel, load_tokenizer
+from numpy.typing import NDArray
+
+load_dotenv()
+
+
+@lru_cache(maxsize=None)
+def load_moonshine(
+    model_name: Literal["moonshine/base", "moonshine/tiny"],
+) -> MoonshineOnnxModel:
+    return MoonshineOnnxModel(model_name=model_name)
+
+
+tokenizer = load_tokenizer()
+
+
+def stt(
+    audio: tuple[int, NDArray[np.int16 | np.float32]],
+    model_name: Literal["moonshine/base", "moonshine/tiny"],
+    captions: str,
+) -> Generator[AdditionalOutputs, None, None]:
+    moonshine = load_moonshine(model_name)
+    sr, audio_np = audio  # type: ignore
+    if audio_np.dtype == np.int16:
+        audio_np = audio_to_float32(audio)
+    if audio_np.ndim == 1:
+        audio_np = audio_np.reshape(1, -1)
+    tokens = moonshine.generate(audio_np)
+    yield AdditionalOutputs(
+        (captions + "\n" + tokenizer.decode_batch(tokens)[0]).strip()
+    )
+
+
+captions = gr.Textbox(label="Captions")
+stream = Stream(
+    ReplyOnPause(stt, input_sample_rate=16000),
+    modality="audio",
+    mode="send",
+    ui_args={
+        "title": "Live Captions by Moonshine",
+        "icon": "default-favicon.ico",
+        "icon_button_color": "#5c5c5c",
+        "pulse_color": "#a7c6fc",
+        "icon_radius": 0,
+    },
+    rtc_configuration=get_twilio_turn_credentials(),
+    additional_inputs=[
+        gr.Radio(
+            choices=["moonshine/base", "moonshine/tiny"],
+            value="moonshine/base",
+            label="Model",
+        ),
+        captions,
+    ],
+    additional_outputs=[captions],
+    additional_outputs_handler=lambda prev, current: (prev + "\n" + current).strip(),
+)
+
+if __name__ == "__main__":
+    stream.ui.launch()

demo/moonshine_live/requirements.txt

@@ -0,0 +1,3 @@
+fastrtc[vad]
+useful-moonshine-onnx@git+https://git@github.com/usefulsensors/moonshine.git#subdirectory=moonshine-onnx
+twilio

demo/nextjs_voice_chat/README.md

@@ -0,0 +1,74 @@
+# FastRTC POC
+A simple POC for a fast real-time voice chat application using FastAPI and FastRTC by [rohanprichard](https://github.com/rohanprichard). I wanted to make one as an example with more production-ready languages, rather than just Gradio.
+
+## Setup
+1. Set your API keys in an `.env` file based on the `.env.example` file
+2. Create a virtual environment and install the dependencies
+    ```bash
+    python3 -m venv env
+    source env/bin/activate
+    pip install -r requirements.txt
+    ```
+
+3. Run the server
+    ```bash
+    ./run.sh
+    ```
+4. Navigate into the frontend directory in another terminal
+    ```bash
+    cd frontend/fastrtc-demo
+    ```
+5. Run the frontend
+    ```bash
+    npm install
+    npm run dev
+    ```
+6. Go to the URL and click the microphone icon to start chatting!
+
+7. Reset chats by clicking the trash button on the bottom right
+
+## Notes
+You can choose to not install the requirements for TTS and STT by removing the `[tts, stt]` from the specifier in the `requirements.txt` file.
+
+- The STT is currently using the ElevenLabs API.
+- The LLM is currently using the OpenAI API.
+- The TTS is currently using the ElevenLabs API.
+- The VAD is currently using the Silero VAD model.
+- You may need to install ffmpeg if you get errors in STT
+
+The prompt can be changed in the `backend/server.py` file and modified as you like.
+
+### Audio Parameters 
+
+#### AlgoOptions
+
+- **audio_chunk_duration**: Length of audio chunks in seconds. Smaller values allow for faster processing but may be less accurate.
+- **started_talking_threshold**: If a chunk has more than this many seconds of speech, the system considers that the user has started talking.
+- **speech_threshold**: After the user has started speaking, if a chunk has less than this many seconds of speech, the system considers that the user has paused.
+
+#### SileroVadOptions
+
+- **threshold**: Speech probability threshold (0.0-1.0). Values above this are considered speech. Higher values are more strict.
+- **min_speech_duration_ms**: Speech segments shorter than this (in milliseconds) are filtered out.
+- **min_silence_duration_ms**: The system waits for this duration of silence (in milliseconds) before considering speech to be finished.
+- **speech_pad_ms**: Padding added to both ends of detected speech segments to prevent cutting off words.
+- **max_speech_duration_s**: Maximum allowed duration for a speech segment in seconds. Prevents indefinite listening.
+
+### Tuning Recommendations
+
+- If the AI interrupts you too early:
+  - Increase `min_silence_duration_ms`
+  - Increase `speech_threshold`
+  - Increase `speech_pad_ms`
+
+- If the AI is slow to respond after you finish speaking:
+  - Decrease `min_silence_duration_ms`
+  - Decrease `speech_threshold`
+
+- If the system fails to detect some speech:
+  - Lower the `threshold` value
+  - Decrease `started_talking_threshold`
+
+
+## Credits:
+Credit for the UI components goes to Shadcn, Aceternity UI and Kokonut UI.

demo/nextjs_voice_chat/backend/env.py

@@ -0,0 +1,7 @@
+from dotenv import load_dotenv
+import os
+
+load_dotenv()
+
+LLM_API_KEY = os.getenv("LLM_API_KEY")
+ELEVENLABS_API_KEY = os.getenv("ELEVENLABS_API_KEY")

demo/nextjs_voice_chat/backend/server.py

@@ -0,0 +1,129 @@
+import fastapi
+from fastrtc import ReplyOnPause, Stream, AlgoOptions, SileroVadOptions
+from fastrtc.utils import audio_to_bytes
+from openai import OpenAI
+import logging
+import time
+from fastapi.middleware.cors import CORSMiddleware
+from elevenlabs import VoiceSettings, stream
+from elevenlabs.client import ElevenLabs
+import numpy as np
+
+from .env import LLM_API_KEY, ELEVENLABS_API_KEY
+
+
+sys_prompt = """
+You are a helpful assistant. You are witty, engaging and fun. You love being interactive with the user. 
+You also can add minimalistic utterances like 'uh-huh' or 'mm-hmm' to the conversation to make it more natural. However, only vocalization are allowed, no actions or other non-vocal sounds.
+Begin a conversation with a self-deprecating joke like 'I'm not sure if I'm ready for this...' or 'I bet you already regret clicking that button...'
+"""
+
+messages = [{"role": "system", "content": sys_prompt}]
+
+openai_client = OpenAI(api_key=LLM_API_KEY)
+
+elevenlabs_client = ElevenLabs(api_key=ELEVENLABS_API_KEY)
+
+logging.basicConfig(level=logging.INFO)
+
+
+def echo(audio):
+    stt_time = time.time()
+
+    logging.info("Performing STT")
+
+    transcription = elevenlabs_client.speech_to_text.convert(
+        file=audio_to_bytes(audio),
+        model_id="scribe_v1",
+        tag_audio_events=False,
+        language_code="eng",
+        diarize=False,
+    )
+    prompt = transcription.text
+    if prompt == "":
+        logging.info("STT returned empty string")
+        return
+    logging.info(f"STT response: {prompt}")
+
+    messages.append({"role": "user", "content": prompt})
+
+    logging.info(f"STT took {time.time() - stt_time} seconds")
+
+    llm_time = time.time()
+
+    def text_stream():
+        global full_response
+        full_response = ""
+
+        response = openai_client.chat.completions.create(
+            model="gpt-3.5-turbo", messages=messages, max_tokens=200, stream=True
+        )
+
+        for chunk in response:
+            if chunk.choices[0].finish_reason == "stop":
+                break
+            if chunk.choices[0].delta.content:
+                full_response += chunk.choices[0].delta.content
+                yield chunk.choices[0].delta.content
+
+    audio_stream = elevenlabs_client.generate(
+        text=text_stream(),
+        voice="Rachel",  # Cassidy is also really good
+        voice_settings=VoiceSettings(
+            similarity_boost=0.9, stability=0.6, style=0.4, speed=1
+        ),
+        model="eleven_multilingual_v2",
+        output_format="pcm_24000",
+        stream=True,
+    )
+
+    for audio_chunk in audio_stream:
+        audio_array = (
+            np.frombuffer(audio_chunk, dtype=np.int16).astype(np.float32) / 32768.0
+        )
+        yield (24000, audio_array)
+
+    messages.append({"role": "assistant", "content": full_response + " "})
+    logging.info(f"LLM response: {full_response}")
+    logging.info(f"LLM took {time.time() - llm_time} seconds")
+
+
+stream = Stream(
+    ReplyOnPause(
+        echo,
+        algo_options=AlgoOptions(
+            audio_chunk_duration=0.5,
+            started_talking_threshold=0.1,
+            speech_threshold=0.03,
+        ),
+        model_options=SileroVadOptions(
+            threshold=0.75,
+            min_speech_duration_ms=250,
+            min_silence_duration_ms=1500,
+            speech_pad_ms=400,
+            max_speech_duration_s=15,
+        ),
+    ),
+    modality="audio",
+    mode="send-receive",
+)
+
+app = fastapi.FastAPI()
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+stream.mount(app)
+
+
+@app.get("/reset")
+async def reset():
+    global messages
+    logging.info("Resetting chat")
+    messages = [{"role": "system", "content": sys_prompt}]
+    return {"status": "success"}

demo/nextjs_voice_chat/frontend/fastrtc-demo/.gitignore

@@ -0,0 +1,41 @@
+# See https://help.github.com/articles/ignoring-files/ for more about ignoring files.
+
+# dependencies
+/node_modules
+/.pnp
+.pnp.*
+.yarn/*
+!.yarn/patches
+!.yarn/plugins
+!.yarn/releases
+!.yarn/versions
+
+# testing
+/coverage
+
+# next.js
+/.next/
+/out/
+
+# production
+/build
+
+# misc
+.DS_Store
+*.pem
+
+# debug
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+.pnpm-debug.log*
+
+# env files (can opt-in for committing if needed)
+.env*
+
+# vercel
+.vercel
+
+# typescript
+*.tsbuildinfo
+next-env.d.ts

demo/nextjs_voice_chat/frontend/fastrtc-demo/README.md

@@ -0,0 +1,36 @@
+This is a [Next.js](https://nextjs.org) project bootstrapped with [`create-next-app`](https://nextjs.org/docs/app/api-reference/cli/create-next-app).
+
+## Getting Started
+
+First, run the development server:
+
+```bash
+npm run dev
+# or
+yarn dev
+# or
+pnpm dev
+# or
+bun dev
+```
+
+Open [http://localhost:3000](http://localhost:3000) with your browser to see the result.
+
+You can start editing the page by modifying `app/page.tsx`. The page auto-updates as you edit the file.
+
+This project uses [`next/font`](https://nextjs.org/docs/app/building-your-application/optimizing/fonts) to automatically optimize and load [Geist](https://vercel.com/font), a new font family for Vercel.
+
+## Learn More
+
+To learn more about Next.js, take a look at the following resources:
+
+- [Next.js Documentation](https://nextjs.org/docs) - learn about Next.js features and API.
+- [Learn Next.js](https://nextjs.org/learn) - an interactive Next.js tutorial.
+
+You can check out [the Next.js GitHub repository](https://github.com/vercel/next.js) - your feedback and contributions are welcome!
+
+## Deploy on Vercel
+
+The easiest way to deploy your Next.js app is to use the [Vercel Platform](https://vercel.com/new?utm_medium=default-template&filter=next.js&utm_source=create-next-app&utm_campaign=create-next-app-readme) from the creators of Next.js.
+
+Check out our [Next.js deployment documentation](https://nextjs.org/docs/app/building-your-application/deploying) for more details.

demo/nextjs_voice_chat/frontend/fastrtc-demo/app/globals.css

@@ -0,0 +1,130 @@
+@import "tailwindcss";
+
+@plugin "tailwindcss-animate";
+
+@custom-variant dark (&:is(.dark *));
+
+@theme inline {
+  --color-background: var(--background);
+  --color-foreground: var(--foreground);
+  --font-sans: var(--font-geist-sans);
+  --font-mono: var(--font-geist-mono);
+  --color-sidebar-ring: var(--sidebar-ring);
+  --color-sidebar-border: var(--sidebar-border);
+  --color-sidebar-accent-foreground: var(--sidebar-accent-foreground);
+  --color-sidebar-accent: var(--sidebar-accent);
+  --color-sidebar-primary-foreground: var(--sidebar-primary-foreground);
+  --color-sidebar-primary: var(--sidebar-primary);
+  --color-sidebar-foreground: var(--sidebar-foreground);
+  --color-sidebar: var(--sidebar);
+  --color-chart-5: var(--chart-5);
+  --color-chart-4: var(--chart-4);
+  --color-chart-3: var(--chart-3);
+  --color-chart-2: var(--chart-2);
+  --color-chart-1: var(--chart-1);
+  --color-ring: var(--ring);
+  --color-input: var(--input);
+  --color-border: var(--border);
+  --color-destructive-foreground: var(--destructive-foreground);
+  --color-destructive: var(--destructive);
+  --color-accent-foreground: var(--accent-foreground);
+  --color-accent: var(--accent);
+  --color-muted-foreground: var(--muted-foreground);
+  --color-muted: var(--muted);
+  --color-secondary-foreground: var(--secondary-foreground);
+  --color-secondary: var(--secondary);
+  --color-primary-foreground: var(--primary-foreground);
+  --color-primary: var(--primary);
+  --color-popover-foreground: var(--popover-foreground);
+  --color-popover: var(--popover);
+  --color-card-foreground: var(--card-foreground);
+  --color-card: var(--card);
+  --radius-sm: calc(var(--radius) - 4px);
+  --radius-md: calc(var(--radius) - 2px);
+  --radius-lg: var(--radius);
+  --radius-xl: calc(var(--radius) + 4px);
+}
+
+:root {
+  --background: oklch(1 0 0);
+  --foreground: oklch(0.129 0.042 264.695);
+  --card: oklch(1 0 0);
+  --card-foreground: oklch(0.129 0.042 264.695);
+  --popover: oklch(1 0 0);
+  --popover-foreground: oklch(0.129 0.042 264.695);
+  --primary: oklch(0.208 0.042 265.755);
+  --primary-foreground: oklch(0.984 0.003 247.858);
+  --secondary: oklch(0.968 0.007 247.896);
+  --secondary-foreground: oklch(0.208 0.042 265.755);
+  --muted: oklch(0.968 0.007 247.896);
+  --muted-foreground: oklch(0.554 0.046 257.417);
+  --accent: oklch(0.968 0.007 247.896);
+  --accent-foreground: oklch(0.208 0.042 265.755);
+  --destructive: oklch(0.577 0.245 27.325);
+  --destructive-foreground: oklch(0.577 0.245 27.325);
+  --border: oklch(0.929 0.013 255.508);
+  --input: oklch(0.929 0.013 255.508);
+  --ring: oklch(0.704 0.04 256.788);
+  --chart-1: oklch(0.646 0.222 41.116);
+  --chart-2: oklch(0.6 0.118 184.704);
+  --chart-3: oklch(0.398 0.07 227.392);
+  --chart-4: oklch(0.828 0.189 84.429);
+  --chart-5: oklch(0.769 0.188 70.08);
+  --radius: 0.625rem;
+  --sidebar: oklch(0.984 0.003 247.858);
+  --sidebar-foreground: oklch(0.129 0.042 264.695);
+  --sidebar-primary: oklch(0.208 0.042 265.755);
+  --sidebar-primary-foreground: oklch(0.984 0.003 247.858);
+  --sidebar-accent: oklch(0.968 0.007 247.896);
+  --sidebar-accent-foreground: oklch(0.208 0.042 265.755);
+  --sidebar-border: oklch(0.929 0.013 255.508);
+  --sidebar-ring: oklch(0.704 0.04 256.788);
+}
+
+.dark {
+  --background: oklch(0.129 0.042 264.695);
+  --foreground: oklch(0.984 0.003 247.858);
+  --card: oklch(0.129 0.042 264.695);
+  --card-foreground: oklch(0.984 0.003 247.858);
+  --popover: oklch(0.129 0.042 264.695);
+  --popover-foreground: oklch(0.984 0.003 247.858);
+  --primary: oklch(0.984 0.003 247.858);
+  --primary-foreground: oklch(0.208 0.042 265.755);
+  --secondary: oklch(0.279 0.041 260.031);
+  --secondary-foreground: oklch(0.984 0.003 247.858);
+  --muted: oklch(0.279 0.041 260.031);
+  --muted-foreground: oklch(0.704 0.04 256.788);
+  --accent: oklch(0.279 0.041 260.031);
+  --accent-foreground: oklch(0.984 0.003 247.858);
+  --destructive: oklch(0.396 0.141 25.723);
+  --destructive-foreground: oklch(0.637 0.237 25.331);
+  --border: oklch(0.279 0.041 260.031);
+  --input: oklch(0.279 0.041 260.031);
+  --ring: oklch(0.446 0.043 257.281);
+  --chart-1: oklch(0.488 0.243 264.376);
+  --chart-2: oklch(0.696 0.17 162.48);
+  --chart-3: oklch(0.769 0.188 70.08);
+  --chart-4: oklch(0.627 0.265 303.9);
+  --chart-5: oklch(0.645 0.246 16.439);
+  --sidebar: oklch(0.208 0.042 265.755);
+  --sidebar-foreground: oklch(0.984 0.003 247.858);
+  --sidebar-primary: oklch(0.488 0.243 264.376);
+  --sidebar-primary-foreground: oklch(0.984 0.003 247.858);
+  --sidebar-accent: oklch(0.279 0.041 260.031);
+  --sidebar-accent-foreground: oklch(0.984 0.003 247.858);
+  --sidebar-border: oklch(0.279 0.041 260.031);
+  --sidebar-ring: oklch(0.446 0.043 257.281);
+}
+
+@layer base {
+  * {
+    @apply border-border outline-ring/50;
+  }
+  body {
+    @apply bg-background text-foreground;
+  }
+}
+
+.no-transitions * {
+  transition: none !important;
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/app/layout.tsx

@@ -0,0 +1,44 @@
+import type { Metadata } from "next";
+import { Geist, Geist_Mono } from "next/font/google";
+import "./globals.css";
+import { ThemeProvider } from "@/components/theme-provider";
+import { ThemeTransition } from "@/components/ui/theme-transition";
+
+const geistSans = Geist({
+  variable: "--font-geist-sans",
+  subsets: ["latin"],
+});
+
+const geistMono = Geist_Mono({
+  variable: "--font-geist-mono",
+  subsets: ["latin"],
+});
+
+export const metadata: Metadata = {
+  title: "FastRTC Demo",
+  description: "Interactive WebRTC demo with audio visualization",
+};
+
+export default function RootLayout({
+  children,
+}: Readonly<{
+  children: React.ReactNode;
+}>) {
+  return (
+    <html lang="en" suppressHydrationWarning>
+      <body
+        className={`${geistSans.variable} ${geistMono.variable} antialiased`}
+      >
+        <ThemeProvider
+          attribute="class"
+          defaultTheme="dark"
+          enableSystem
+          disableTransitionOnChange
+        >
+          {children}
+          <ThemeTransition />
+        </ThemeProvider>
+      </body>
+    </html>
+  );
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/app/page.tsx

@@ -0,0 +1,16 @@
+import { BackgroundCircleProvider } from "@/components/background-circle-provider";
+import { ThemeToggle } from "@/components/ui/theme-toggle";
+import { ResetChat } from "@/components/ui/reset-chat";
+export default function Home() {
+  return (
+    <div className="flex flex-col items-center justify-center h-screen">
+      <BackgroundCircleProvider />
+      <div className="absolute top-4 right-4 z-10">
+        <ThemeToggle />
+      </div>
+      <div className="absolute bottom-4 right-4 z-10">
+        <ResetChat />
+      </div>
+    </div>
+  );
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/components.json

@@ -0,0 +1,21 @@
+{
+  "$schema": "https://ui.shadcn.com/schema.json",
+  "style": "new-york",
+  "rsc": true,
+  "tsx": true,
+  "tailwind": {
+    "config": "",
+    "css": "app/globals.css",
+    "baseColor": "slate",
+    "cssVariables": true,
+    "prefix": ""
+  },
+  "aliases": {
+    "components": "@/components",
+    "utils": "@/lib/utils",
+    "ui": "@/components/ui",
+    "lib": "@/lib",
+    "hooks": "@/hooks"
+  },
+  "iconLibrary": "lucide"
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/background-circle-provider.tsx

@@ -0,0 +1,123 @@
+"use client"
+
+import { useState, useEffect, useRef, useCallback } from "react";
+import { BackgroundCircles } from "@/components/ui/background-circles";
+import { AIVoiceInput } from "@/components/ui/ai-voice-input";
+import { WebRTCClient } from "@/lib/webrtc-client";
+
+export function BackgroundCircleProvider() {
+    const [currentVariant, setCurrentVariant] = 
+        useState<keyof typeof COLOR_VARIANTS>("octonary");
+    const [isConnected, setIsConnected] = useState(false);
+    const [webrtcClient, setWebrtcClient] = useState<WebRTCClient | null>(null);
+    const [audioLevel, setAudioLevel] = useState(0);
+    const audioRef = useRef<HTMLAudioElement>(null);
+
+    // Memoize callbacks to prevent recreation on each render
+    const handleConnected = useCallback(() => setIsConnected(true), []);
+    const handleDisconnected = useCallback(() => setIsConnected(false), []);
+    
+    const handleAudioStream = useCallback((stream: MediaStream) => {
+        if (audioRef.current) {
+            audioRef.current.srcObject = stream;
+        }
+    }, []);
+    
+    const handleAudioLevel = useCallback((level: number) => {
+        // Apply some smoothing to the audio level
+        setAudioLevel(prev => prev * 0.7 + level * 0.3);
+    }, []);
+
+    // Get all available variants
+    const variants = Object.keys(
+        COLOR_VARIANTS
+    ) as (keyof typeof COLOR_VARIANTS)[];
+
+    // Function to change to the next color variant
+    const changeVariant = () => {
+        const currentIndex = variants.indexOf(currentVariant);
+        const nextVariant = variants[(currentIndex + 1) % variants.length];
+        setCurrentVariant(nextVariant);
+    };
+
+    useEffect(() => {
+        // Initialize WebRTC client with memoized callbacks
+        const client = new WebRTCClient({
+            onConnected: handleConnected,
+            onDisconnected: handleDisconnected,
+            onAudioStream: handleAudioStream,
+            onAudioLevel: handleAudioLevel
+        });
+        setWebrtcClient(client);
+
+        return () => {
+            client.disconnect();
+        };
+    }, [handleConnected, handleDisconnected, handleAudioStream, handleAudioLevel]);
+
+    const handleStart = () => {
+        webrtcClient?.connect();
+    };
+
+    const handleStop = () => {
+        webrtcClient?.disconnect();
+    };
+
+    return (
+        <div 
+            className="relative w-full h-full"
+            onClick={changeVariant} // Add click handler to change color
+        >
+            <BackgroundCircles 
+                variant={currentVariant} 
+                audioLevel={audioLevel}
+                isActive={isConnected}
+            />
+            <div className="absolute inset-0 flex items-center justify-center">
+                <AIVoiceInput 
+                    onStart={handleStart}
+                    onStop={handleStop}
+                    isConnected={isConnected}
+                />
+            </div>
+            <audio ref={audioRef} autoPlay hidden />
+        </div>
+    );
+}
+
+export default { BackgroundCircleProvider }
+
+const COLOR_VARIANTS = {
+    primary: {
+        border: [
+            "border-emerald-500/60",
+            "border-cyan-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-emerald-500/30",
+    },
+    secondary: {
+        border: [
+            "border-violet-500/60",
+            "border-fuchsia-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-violet-500/30",
+    },
+    senary: {
+        border: [
+            "border-blue-500/60",
+            "border-sky-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-blue-500/30",
+    }, // blue
+    octonary: {
+        border: [
+            "border-red-500/60",
+            "border-rose-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-red-500/30",
+    },
+} as const;

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/theme-provider.tsx

@@ -0,0 +1,101 @@
+"use client";
+
+import { createContext, useContext, useEffect, useState } from "react";
+
+type Theme = "light" | "dark" | "system";
+
+type ThemeProviderProps = {
+  children: React.ReactNode;
+  defaultTheme?: Theme;
+  storageKey?: string;
+  attribute?: string;
+  enableSystem?: boolean;
+  disableTransitionOnChange?: boolean;
+};
+
+type ThemeProviderState = {
+  theme: Theme;
+  setTheme: (theme: Theme) => void;
+};
+
+const initialState: ThemeProviderState = {
+  theme: "system",
+  setTheme: () => null,
+};
+
+const ThemeProviderContext = createContext<ThemeProviderState>(initialState);
+
+export function ThemeProvider({
+  children,
+  defaultTheme = "system",
+  storageKey = "theme",
+  attribute = "class",
+  enableSystem = true,
+  disableTransitionOnChange = false,
+  ...props
+}: ThemeProviderProps) {
+  const [theme, setTheme] = useState<Theme>(defaultTheme);
+
+  useEffect(() => {
+    const savedTheme = localStorage.getItem(storageKey) as Theme | null;
+    
+    if (savedTheme) {
+      setTheme(savedTheme);
+    } else if (defaultTheme === "system" && enableSystem) {
+      const systemTheme = window.matchMedia("(prefers-color-scheme: dark)").matches
+        ? "dark"
+        : "light";
+      setTheme(systemTheme);
+    }
+  }, [defaultTheme, storageKey, enableSystem]);
+
+  useEffect(() => {
+    const root = window.document.documentElement;
+    
+    if (disableTransitionOnChange) {
+      root.classList.add("no-transitions");
+      
+      // Force a reflow
+      window.getComputedStyle(root).getPropertyValue("opacity");
+      
+      setTimeout(() => {
+        root.classList.remove("no-transitions");
+      }, 0);
+    }
+    
+    root.classList.remove("light", "dark");
+    
+    if (theme === "system" && enableSystem) {
+      const systemTheme = window.matchMedia("(prefers-color-scheme: dark)").matches
+        ? "dark"
+        : "light";
+      root.classList.add(systemTheme);
+    } else {
+      root.classList.add(theme);
+    }
+
+    localStorage.setItem(storageKey, theme);
+  }, [theme, storageKey, enableSystem, disableTransitionOnChange]);
+
+  const value = {
+    theme,
+    setTheme: (theme: Theme) => {
+      setTheme(theme);
+    },
+  };
+
+  return (
+    <ThemeProviderContext.Provider {...props} value={value}>
+      {children}
+    </ThemeProviderContext.Provider>
+  );
+}
+
+export const useTheme = () => {
+  const context = useContext(ThemeProviderContext);
+
+  if (context === undefined)
+    throw new Error("useTheme must be used within a ThemeProvider");
+
+  return context;
+};

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/ui/ai-voice-input.tsx

@@ -0,0 +1,114 @@
+"use client";
+
+import { Mic, Square } from "lucide-react";
+import { useState, useEffect } from "react";
+import { cn } from "@/lib/utils";
+
+interface AIVoiceInputProps {
+  onStart?: () => void;
+  onStop?: (duration: number) => void;
+  isConnected?: boolean;
+  className?: string;
+}
+
+export function AIVoiceInput({
+  onStart,
+  onStop,
+  isConnected = false,
+  className
+}: AIVoiceInputProps) {
+  const [active, setActive] = useState(false);
+  const [time, setTime] = useState(0);
+  const [isClient, setIsClient] = useState(false);
+  const [status, setStatus] = useState<'disconnected' | 'connecting' | 'connected'>('disconnected');
+
+  useEffect(() => {
+    setIsClient(true);
+  }, []);
+
+  useEffect(() => {
+    let intervalId: NodeJS.Timeout;
+
+    if (active) {
+      intervalId = setInterval(() => {
+        setTime((t) => t + 1);
+      }, 1000);
+    } else {
+      setTime(0);
+    }
+
+    return () => clearInterval(intervalId);
+  }, [active]);
+
+  useEffect(() => {
+    if (isConnected) {
+      setStatus('connected');
+      setActive(true);
+    } else {
+      setStatus('disconnected');
+      setActive(false);
+    }
+  }, [isConnected]);
+
+  const formatTime = (seconds: number) => {
+    const mins = Math.floor(seconds / 60);
+    const secs = seconds % 60;
+    return `${mins.toString().padStart(2, "0")}:${secs.toString().padStart(2, "0")}`;
+  };
+
+  const handleStart = () => {
+    setStatus('connecting');
+    onStart?.();
+  };
+
+  const handleStop = () => {
+    onStop?.(time);
+    setStatus('disconnected');
+  };
+
+  return (
+    <div className={cn("w-full py-4", className)}>
+      <div className="relative max-w-xl w-full mx-auto flex items-center flex-col gap-4">
+        <div className={cn(
+          "px-2 py-1 rounded-md text-xs font-medium bg-black/10 dark:bg-white/10 text-gray-700 dark:text-white"
+        )}>
+          {status === 'connected' ? 'Connected' : status === 'connecting' ? 'Connecting...' : 'Disconnected'}
+        </div>
+
+        <button
+          className={cn(
+            "group w-16 h-16 rounded-xl flex items-center justify-center transition-colors",
+            active
+              ? "bg-red-500/20 hover:bg-red-500/30"
+              : "bg-black/10 hover:bg-black/20 dark:bg-white/10 dark:hover:bg-white/20"
+          )}
+          type="button"
+          onClick={active ? handleStop : handleStart}
+          disabled={status === 'connecting'}
+        >
+          {status === 'connecting' ? (
+            <div
+              className="w-6 h-6 rounded-sm animate-spin bg-black dark:bg-white cursor-pointer pointer-events-auto"
+              style={{ animationDuration: "3s" }}
+            />
+          ) : active ? (
+            <Square className="w-6 h-6 text-red-500" />
+          ) : (
+            <Mic className="w-6 h-6 text-black/70 dark:text-white/70" />
+          )}
+        </button>
+
+        <span
+          className={cn(
+            "font-mono text-sm transition-opacity duration-300",
+            active
+              ? "text-black/70 dark:text-white/70"
+              : "text-black/30 dark:text-white/30"
+          )}
+        >
+          {formatTime(time)}
+        </span>
+      </div>
+    </div>
+  );
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/ui/background-circles.tsx

@@ -0,0 +1,309 @@
+"use client";
+
+import { motion } from "framer-motion";
+import clsx from "clsx";
+import { useState, useEffect } from "react";
+
+interface BackgroundCirclesProps {
+    title?: string;
+    description?: string;
+    className?: string;
+    variant?: keyof typeof COLOR_VARIANTS;
+    audioLevel?: number;
+    isActive?: boolean;
+}
+
+const COLOR_VARIANTS = {
+    primary: {
+        border: [
+            "border-emerald-500/60",
+            "border-cyan-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-emerald-500/30",
+    },
+    secondary: {
+        border: [
+            "border-violet-500/60",
+            "border-fuchsia-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-violet-500/30",
+    },
+    tertiary: {
+        border: [
+            "border-orange-500/60",
+            "border-yellow-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-orange-500/30",
+    },
+    quaternary: {
+        border: [
+            "border-purple-500/60",
+            "border-pink-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-purple-500/30",
+    },
+    quinary: {
+        border: [
+            "border-red-500/60",
+            "border-rose-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-red-500/30",
+    }, // red
+    senary: {
+        border: [
+            "border-blue-500/60",
+            "border-sky-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-blue-500/30",
+    }, // blue
+    septenary: {
+        border: [
+            "border-gray-500/60",
+            "border-gray-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-gray-500/30",
+    },
+    octonary: {
+        border: [
+            "border-red-500/60",
+            "border-rose-400/50",
+            "border-slate-600/30",
+        ],
+        gradient: "from-red-500/30",
+    },
+} as const;
+
+const AnimatedGrid = () => (
+    <motion.div
+        className="absolute inset-0 [mask-image:radial-gradient(ellipse_at_center,transparent_30%,black)]"
+        animate={{
+            backgroundPosition: ["0% 0%", "100% 100%"],
+        }}
+        transition={{
+            duration: 40,
+            repeat: Number.POSITIVE_INFINITY,
+            ease: "linear",
+        }}
+    >
+        <div className="h-full w-full [background-image:repeating-linear-gradient(100deg,#64748B_0%,#64748B_1px,transparent_1px,transparent_4%)] opacity-20" />
+    </motion.div>
+);
+
+export function BackgroundCircles({
+    title = "",
+    description = "",
+    className,
+    variant = "octonary",
+    audioLevel = 0,
+    isActive = false,
+}: BackgroundCirclesProps) {
+    const variantStyles = COLOR_VARIANTS[variant];
+    const [animationParams, setAnimationParams] = useState({
+        scale: 1,
+        duration: 5,
+        intensity: 0
+    });
+    const [isLoaded, setIsLoaded] = useState(false);
+    
+    // Initial page load animation
+    useEffect(() => {
+        // Small delay to ensure the black screen is visible first
+        const timer = setTimeout(() => {
+            setIsLoaded(true);
+        }, 300);
+        
+        return () => clearTimeout(timer);
+    }, []);
+    
+    // Update animation based on audio level
+    useEffect(() => {
+        if (isActive && audioLevel > 0) {
+            // Simple enhancement of audio level for more dramatic effect
+            const enhancedLevel = Math.min(1, audioLevel * 1.5);
+            
+            setAnimationParams({
+                scale: 1 + enhancedLevel * 0.3,
+                duration: Math.max(2, 5 - enhancedLevel * 3),
+                intensity: enhancedLevel
+            });
+        } else if (animationParams.intensity > 0) {
+            // Only reset if we need to (prevents unnecessary updates)
+            const timer = setTimeout(() => {
+                setAnimationParams({
+                    scale: 1,
+                    duration: 5,
+                    intensity: 0
+                });
+            }, 300);
+            
+            return () => clearTimeout(timer);
+        }
+    }, [audioLevel, isActive, animationParams.intensity]);
+    
+    return (
+        <>
+            {/* Initial black overlay that fades out */}
+            <motion.div 
+                className="fixed inset-0 bg-black z-50"
+                initial={{ opacity: 1 }}
+                animate={{ opacity: isLoaded ? 0 : 1 }}
+                transition={{ duration: 1.2, ease: "easeInOut" }}
+                style={{ pointerEvents: isLoaded ? "none" : "auto" }}
+            />
+            
+            <div
+                className={clsx(
+                    "relative flex h-screen w-full items-center justify-center overflow-hidden",
+                    "bg-white dark:bg-black/5",
+                    className
+                )}
+            >
+                <AnimatedGrid />
+                <motion.div 
+                    className="absolute h-[480px] w-[480px]"
+                    initial={{ opacity: 0, scale: 0.9 }}
+                    animate={{ 
+                        opacity: isLoaded ? 1 : 0,
+                        scale: isLoaded ? 1 : 0.9
+                    }}
+                    transition={{ 
+                        duration: 1.5, 
+                        delay: 0.3,
+                        ease: "easeOut" 
+                    }}
+                >
+                    {[0, 1, 2].map((i) => (
+                        <motion.div
+                            key={i}
+                            className={clsx(
+                                "absolute inset-0 rounded-full",
+                                "border-2 bg-gradient-to-br to-transparent",
+                                variantStyles.border[i],
+                                variantStyles.gradient
+                            )}
+                            animate={{
+                                rotate: 360,
+                                scale: [
+                                    1 + (i * 0.05), 
+                                    (1 + (i * 0.05)) * (1 + (isActive ? animationParams.intensity * 0.2 : 0.02)),
+                                    1 + (i * 0.05)
+                                ],
+                                opacity: [
+                                    0.7 + (i * 0.1),
+                                    0.8 + (i * 0.1) + (isActive ? animationParams.intensity * 0.2 : 0),
+                                    0.7 + (i * 0.1)
+                                ]
+                            }}
+                            transition={{
+                                duration: isActive ? animationParams.duration : 8 + (i * 2),
+                                repeat: Number.POSITIVE_INFINITY,
+                                ease: "easeInOut",
+                            }}
+                        >
+                            <div
+                                className={clsx(
+                                    "absolute inset-0 rounded-full mix-blend-screen",
+                                    `bg-[radial-gradient(ellipse_at_center,${variantStyles.gradient.replace(
+                                        "from-",
+                                        ""
+                                    )}/10%,transparent_70%)]`
+                                )}
+                            />
+                        </motion.div>
+                    ))}
+                </motion.div>
+
+                <div className="absolute inset-0 [mask-image:radial-gradient(90%_60%_at_50%_50%,#000_40%,transparent)]">
+                    <motion.div 
+                        className="absolute inset-0 bg-[radial-gradient(ellipse_at_center,#0F766E/30%,transparent_70%)] blur-[120px]"
+                        initial={{ opacity: 0 }}
+                        animate={{
+                            opacity: isLoaded ? 0.7 : 0,
+                            scale: [1, 1 + (isActive ? animationParams.intensity * 0.3 : 0.02), 1],
+                        }}
+                        transition={{
+                            opacity: { duration: 1.8, delay: 0.5 },
+                            scale: { 
+                                duration: isActive ? 2 : 12,
+                                repeat: Number.POSITIVE_INFINITY,
+                                ease: "easeInOut",
+                            }
+                        }}
+                    />
+                    <motion.div 
+                        className="absolute inset-0 bg-[radial-gradient(ellipse_at_center,#2DD4BF/15%,transparent)] blur-[80px]"
+                        initial={{ opacity: 0 }}
+                        animate={{
+                            opacity: isLoaded ? 1 : 0,
+                            scale: [1, 1 + (isActive ? animationParams.intensity * 0.4 : 0.03), 1]
+                        }}
+                        transition={{
+                            opacity: { duration: 2, delay: 0.7 },
+                            scale: { 
+                                duration: isActive ? 1.5 : 15,
+                                repeat: Number.POSITIVE_INFINITY,
+                                ease: "easeInOut",
+                            }
+                        }}
+                    />
+                    
+                    {/* Additional glow that appears only during high audio levels */}
+                    {isActive && animationParams.intensity > 0.4 && (
+                        <motion.div 
+                            className={`absolute inset-0 bg-[radial-gradient(ellipse_at_center,${variantStyles.gradient.replace("from-", "")}/20%,transparent_70%)] blur-[60px]`}
+                            initial={{ opacity: 0, scale: 0.8 }}
+                            animate={{
+                                opacity: [0, animationParams.intensity * 0.6, 0],
+                                scale: [0.8, 1.1, 0.8],
+                            }}
+                            transition={{
+                                duration: 0.8,
+                                repeat: Number.POSITIVE_INFINITY,
+                                ease: "easeInOut",
+                            }}
+                        />
+                    )}
+                </div>
+            </div>
+        </>
+    );
+}
+
+export function DemoCircles() {
+    const [currentVariant, setCurrentVariant] =
+        useState<keyof typeof COLOR_VARIANTS>("octonary");
+
+    const variants = Object.keys(
+        COLOR_VARIANTS
+    ) as (keyof typeof COLOR_VARIANTS)[];
+
+    function getNextVariant() {
+        const currentIndex = variants.indexOf(currentVariant);
+        const nextVariant = variants[(currentIndex + 1) % variants.length];
+        return nextVariant;
+    }
+
+    return (
+        <>
+            <BackgroundCircles variant={currentVariant} />
+            <div className="absolute top-12 right-12">
+                <button
+                    type="button"
+                    className="bg-slate-950 dark:bg-white text-white dark:text-slate-950 px-4 py-1 rounded-md z-10 text-sm font-medium"
+                    onClick={() => {
+                        setCurrentVariant(getNextVariant());
+                    }}
+                >
+                    Change Variant
+                </button>
+            </div>
+        </>
+    );
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/ui/reset-chat.tsx

@@ -0,0 +1,18 @@
+"use client"
+
+import { Trash } from "lucide-react"
+
+export function ResetChat() {
+    return (
+        <button
+            className="w-10 h-10 rounded-md flex items-center justify-center transition-colors relative overflow-hidden bg-black/10 hover:bg-black/20 dark:bg-white/10 dark:hover:bg-white/20"
+            aria-label="Reset chat"
+            onClick={() => fetch("http://localhost:8000/reset")}
+        >
+            <div className="relative z-10">
+                <Trash className="h-5 w-5 text-black/70 dark:text-white/70" />
+            </div>
+    </button>
+    )
+}
+

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/ui/theme-toggle.tsx

@@ -0,0 +1,61 @@
+"use client";
+
+import { useTheme } from "@/components/theme-provider";
+import { cn } from "@/lib/utils";
+import { Moon, Sun } from "lucide-react";
+import { useRef } from "react";
+
+interface ThemeToggleProps {
+  className?: string;
+}
+
+export function ThemeToggle({ className }: ThemeToggleProps) {
+  const { theme } = useTheme();
+  const buttonRef = useRef<HTMLButtonElement>(null);
+
+  const toggleTheme = () => {
+    // Instead of directly changing the theme, dispatch a custom event
+    const newTheme = theme === "light" ? "dark" : "light";
+    
+    // Dispatch custom event with the new theme
+    window.dispatchEvent(
+      new CustomEvent('themeToggleRequest', { 
+        detail: { theme: newTheme } 
+      })
+    );
+  };
+
+  return (
+    <button
+      ref={buttonRef}
+      onClick={toggleTheme}
+      className={cn(
+        "w-10 h-10 rounded-md flex items-center justify-center transition-colors relative overflow-hidden",
+        "bg-black/10 hover:bg-black/20 dark:bg-white/10 dark:hover:bg-white/20",
+        className
+      )}
+      aria-label="Toggle theme"
+    >
+      <div className="relative z-10">
+        {theme === "light" ? (
+          <Moon className="h-5 w-5 text-black/70" />
+        ) : (
+          <Sun className="h-5 w-5 text-white/70" />
+        )}
+      </div>
+      
+      {/* Small inner animation for the button itself */}
+      <div 
+        className={cn(
+          "absolute inset-0 transition-transform duration-500",
+          theme === "light" 
+            ? "bg-gradient-to-br from-blue-500/20 to-purple-500/20 translate-y-full" 
+            : "bg-gradient-to-br from-amber-500/20 to-orange-500/20 -translate-y-full"
+        )}
+        style={{
+          transitionTimingFunction: "cubic-bezier(0.22, 1, 0.36, 1)"
+        }}
+      />
+    </button>
+  );
+} 

demo/nextjs_voice_chat/frontend/fastrtc-demo/components/ui/theme-transition.tsx

@@ -0,0 +1,120 @@
+"use client";
+
+import { useTheme } from "@/components/theme-provider";
+import { useEffect, useState } from "react";
+import { motion, AnimatePresence } from "framer-motion";
+
+interface ThemeTransitionProps {
+  className?: string;
+}
+
+export function ThemeTransition({ className }: ThemeTransitionProps) {
+  const { theme, setTheme } = useTheme();
+  const [position, setPosition] = useState({ x: 0, y: 0 });
+  const [isAnimating, setIsAnimating] = useState(false);
+  const [pendingTheme, setPendingTheme] = useState<string | null>(null);
+  const [visualTheme, setVisualTheme] = useState<string | null>(theme);
+
+  // Track mouse/touch position for click events
+  useEffect(() => {
+    const handleMouseMove = (e: MouseEvent) => {
+      setPosition({ x: e.clientX, y: e.clientY });
+    };
+    
+    const handleTouchMove = (e: TouchEvent) => {
+      if (e.touches[0]) {
+        setPosition({ x: e.touches[0].clientX, y: e.touches[0].clientY });
+      }
+    };
+    
+    window.addEventListener("mousemove", handleMouseMove);
+    window.addEventListener("touchmove", handleTouchMove);
+    
+    return () => {
+      window.removeEventListener("mousemove", handleMouseMove);
+      window.removeEventListener("touchmove", handleTouchMove);
+    };
+  }, []);
+
+  // Listen for theme toggle requests
+  useEffect(() => {
+    // Custom event for theme toggle requests
+    const handleThemeToggle = (e: CustomEvent) => {
+      if (isAnimating) return; // Prevent multiple animations
+      
+      const newTheme = e.detail.theme;
+      if (newTheme === theme) return;
+      
+      // Store the pending theme but don't apply it yet
+      setPendingTheme(newTheme);
+      setIsAnimating(true);
+      
+      // The actual theme will be applied mid-animation
+    };
+
+    window.addEventListener('themeToggleRequest' as any, handleThemeToggle as EventListener);
+    
+    return () => {
+      window.removeEventListener('themeToggleRequest' as any, handleThemeToggle as EventListener);
+    };
+  }, [theme, isAnimating]);
+
+  // Apply the theme change mid-animation
+  useEffect(() => {
+    if (isAnimating && pendingTheme) {
+      // Set visual theme immediately for the animation
+      setVisualTheme(pendingTheme);
+      
+      // Apply the actual theme change after a delay (mid-animation)
+      const timer = setTimeout(() => {
+        setTheme(pendingTheme as any);
+      }, 400); // Half of the animation duration
+      
+      // End the animation after it completes
+      const endTimer = setTimeout(() => {
+        setIsAnimating(false);
+        setPendingTheme(null);
+      }, 1000); // Match with animation duration
+      
+      return () => {
+        clearTimeout(timer);
+        clearTimeout(endTimer);
+      };
+    }
+  }, [isAnimating, pendingTheme, setTheme]);
+
+  return (
+    <AnimatePresence>
+      {isAnimating && (
+        <motion.div
+          className="fixed inset-0 z-[9999] pointer-events-none"
+          initial={{ opacity: 0 }}
+          animate={{ opacity: 1 }}
+          exit={{ opacity: 0 }}
+          transition={{ duration: 0.3 }}
+        >
+          <motion.div
+            className={`absolute rounded-full ${visualTheme === 'dark' ? 'bg-slate-950' : 'bg-white'}`}
+            initial={{ 
+              width: 0, 
+              height: 0,
+              x: position.x,
+              y: position.y,
+              borderRadius: '100%' 
+            }}
+            animate={{ 
+              width: Math.max(window.innerWidth * 3, window.innerHeight * 3),
+              height: Math.max(window.innerWidth * 3, window.innerHeight * 3),
+              x: position.x - Math.max(window.innerWidth * 3, window.innerHeight * 3) / 2,
+              y: position.y - Math.max(window.innerWidth * 3, window.innerHeight * 3) / 2,
+            }}
+            transition={{ 
+              duration: 0.8,
+              ease: [0.22, 1, 0.36, 1]
+            }}
+          />
+        </motion.div>
+      )}
+    </AnimatePresence>
+  );
+} 

demo/nextjs_voice_chat/frontend/fastrtc-demo/eslint.config.mjs

@@ -0,0 +1,28 @@
+import { dirname } from "path";
+import { fileURLToPath } from "url";
+import { FlatCompat } from "@eslint/eslintrc";
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = dirname(__filename);
+
+const compat = new FlatCompat({
+  baseDirectory: __dirname,
+});
+
+const eslintConfig = [
+  ...compat.extends("next/core-web-vitals", "next/typescript"),
+  {
+    rules: {
+      "no-unused-vars": "off",
+      "no-explicit-any": "off",
+      "no-console": "off",
+      "no-debugger": "off",
+      "eqeqeq": "off",
+      "curly": "off",
+      "quotes": "off",
+      "semi": "off",
+    },
+  },
+];
+
+export default eslintConfig;

demo/nextjs_voice_chat/frontend/fastrtc-demo/lib/utils.ts

@@ -0,0 +1,6 @@
+import { clsx, type ClassValue } from "clsx"
+import { twMerge } from "tailwind-merge"
+
+export function cn(...inputs: ClassValue[]) {
+  return twMerge(clsx(inputs))
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/lib/webrtc-client.ts

@@ -0,0 +1,189 @@
+interface WebRTCClientOptions {
+    onConnected?: () => void;
+    onDisconnected?: () => void;
+    onMessage?: (message: any) => void;
+    onAudioStream?: (stream: MediaStream) => void;
+    onAudioLevel?: (level: number) => void;
+}
+
+export class WebRTCClient {
+    private peerConnection: RTCPeerConnection | null = null;
+    private mediaStream: MediaStream | null = null;
+    private dataChannel: RTCDataChannel | null = null;
+    private options: WebRTCClientOptions;
+    private audioContext: AudioContext | null = null;
+    private analyser: AnalyserNode | null = null;
+    private dataArray: Uint8Array | null = null;
+    private animationFrameId: number | null = null;
+
+    constructor(options: WebRTCClientOptions = {}) {
+        this.options = options;
+    }
+
+    async connect() {
+        try {
+            this.peerConnection = new RTCPeerConnection();
+            
+            // Get user media
+            try {
+                this.mediaStream = await navigator.mediaDevices.getUserMedia({
+                    audio: true
+                });
+            } catch (mediaError: any) {
+                console.error('Media error:', mediaError);
+                if (mediaError.name === 'NotAllowedError') {
+                    throw new Error('Microphone access denied. Please allow microphone access and try again.');
+                } else if (mediaError.name === 'NotFoundError') {
+                    throw new Error('No microphone detected. Please connect a microphone and try again.');
+                } else {
+                    throw mediaError;
+                }
+            }
+            
+            this.setupAudioAnalysis();
+            
+            this.mediaStream.getTracks().forEach(track => {
+                if (this.peerConnection) {
+                    this.peerConnection.addTrack(track, this.mediaStream!);
+                }
+            });
+            
+            this.peerConnection.addEventListener('track', (event) => {
+                if (this.options.onAudioStream) {
+                    this.options.onAudioStream(event.streams[0]);
+                }
+            });
+            
+            this.dataChannel = this.peerConnection.createDataChannel('text');
+            
+            this.dataChannel.addEventListener('message', (event) => {
+                try {
+                    const message = JSON.parse(event.data);
+                    console.log('Received message:', message);
+                    
+                    if (this.options.onMessage) {
+                        this.options.onMessage(message);
+                    }
+                } catch (error) {
+                    console.error('Error parsing message:', error);
+                }
+            });
+            
+            // Create and send offer
+            const offer = await this.peerConnection.createOffer();
+            await this.peerConnection.setLocalDescription(offer);
+            
+            // Use same-origin request to avoid CORS preflight
+            const response = await fetch('http://localhost:8000/webrtc/offer', {
+                method: 'POST',
+                headers: { 
+                    'Content-Type': 'application/json',
+                    'Accept': 'application/json'
+                },
+                mode: 'cors', // Explicitly set CORS mode
+                credentials: 'same-origin',
+                body: JSON.stringify({
+                    sdp: offer.sdp,
+                    type: offer.type,
+                    webrtc_id: Math.random().toString(36).substring(7)
+                })
+            });
+            
+            const serverResponse = await response.json();
+            await this.peerConnection.setRemoteDescription(serverResponse);
+            
+            if (this.options.onConnected) {
+                this.options.onConnected();
+            }
+        } catch (error) {
+            console.error('Error connecting:', error);
+            this.disconnect();
+            throw error;
+        }
+    }
+
+    private setupAudioAnalysis() {
+        if (!this.mediaStream) return;
+        
+        try {
+            this.audioContext = new AudioContext();
+            this.analyser = this.audioContext.createAnalyser();
+            this.analyser.fftSize = 256;
+            
+            const source = this.audioContext.createMediaStreamSource(this.mediaStream);
+            source.connect(this.analyser);
+            
+            const bufferLength = this.analyser.frequencyBinCount;
+            this.dataArray = new Uint8Array(bufferLength);
+            
+            this.startAnalysis();
+        } catch (error) {
+            console.error('Error setting up audio analysis:', error);
+        }
+    }
+
+    private startAnalysis() {
+        if (!this.analyser || !this.dataArray || !this.options.onAudioLevel) return;
+        
+        // Add throttling to prevent too many updates
+        let lastUpdateTime = 0;
+        const throttleInterval = 100; // Only update every 100ms
+        
+        const analyze = () => {
+            this.analyser!.getByteFrequencyData(this.dataArray!);
+            
+            const currentTime = Date.now();
+            // Only update if enough time has passed since last update
+            if (currentTime - lastUpdateTime > throttleInterval) {
+                // Calculate average volume level (0-1)
+                let sum = 0;
+                for (let i = 0; i < this.dataArray!.length; i++) {
+                    sum += this.dataArray![i];
+                }
+                const average = sum / this.dataArray!.length / 255;
+                
+                this.options.onAudioLevel!(average);
+                lastUpdateTime = currentTime;
+            }
+            
+            this.animationFrameId = requestAnimationFrame(analyze);
+        };
+        
+        this.animationFrameId = requestAnimationFrame(analyze);
+    }
+
+    private stopAnalysis() {
+        if (this.animationFrameId !== null) {
+            cancelAnimationFrame(this.animationFrameId);
+            this.animationFrameId = null;
+        }
+        
+        if (this.audioContext) {
+            this.audioContext.close();
+            this.audioContext = null;
+        }
+        
+        this.analyser = null;
+        this.dataArray = null;
+    }
+
+    disconnect() {
+        this.stopAnalysis();
+        
+        if (this.mediaStream) {
+            this.mediaStream.getTracks().forEach(track => track.stop());
+            this.mediaStream = null;
+        }
+        
+        if (this.peerConnection) {
+            this.peerConnection.close();
+            this.peerConnection = null;
+        }
+        
+        this.dataChannel = null;
+        
+        if (this.options.onDisconnected) {
+            this.options.onDisconnected();
+        }
+    }
+} 

demo/nextjs_voice_chat/frontend/fastrtc-demo/next.config.ts

@@ -0,0 +1,7 @@
+import type { NextConfig } from "next";
+
+const nextConfig: NextConfig = {
+  /* config options here */
+};
+
+export default nextConfig;

demo/nextjs_voice_chat/frontend/fastrtc-demo/package.json

@@ -0,0 +1,33 @@
+{
+  "name": "fastrtc-demo",
+  "version": "0.1.0",
+  "private": true,
+  "scripts": {
+    "dev": "next dev --turbopack",
+    "build": "next build --no-lint",
+    "start": "next start",
+    "lint": "next lint"
+  },
+  "dependencies": {
+    "class-variance-authority": "^0.7.1",
+    "clsx": "^2.1.1",
+    "framer-motion": "^12.4.10",
+    "lucide-react": "^0.477.0",
+    "next": "15.2.2-canary.1",
+    "react": "^19.0.0",
+    "react-dom": "^19.0.0",
+    "tailwind-merge": "^3.0.2",
+    "tailwindcss-animate": "^1.0.7"
+  },
+  "devDependencies": {
+    "@eslint/eslintrc": "^3",
+    "@tailwindcss/postcss": "^4",
+    "@types/node": "^20",
+    "@types/react": "^19",
+    "@types/react-dom": "^19",
+    "eslint": "^9",
+    "eslint-config-next": "15.2.2-canary.1",
+    "tailwindcss": "^4",
+    "typescript": "^5"
+  }
+}

demo/nextjs_voice_chat/frontend/fastrtc-demo/postcss.config.mjs

@@ -0,0 +1,5 @@
+const config = {
+  plugins: ["@tailwindcss/postcss"],
+};
+
+export default config;

demo/nextjs_voice_chat/frontend/fastrtc-demo/public/file.svg

@@ -0,0 +1 @@
+<svg fill="none" viewBox="0 0 16 16" xmlns="http://www.w3.org/2000/svg"><path d="M14.5 13.5V5.41a1 1 0 0 0-.3-.7L9.8.29A1 1 0 0 0 9.08 0H1.5v13.5A2.5 2.5 0 0 0 4 16h8a2.5 2.5 0 0 0 2.5-2.5m-1.5 0v-7H8v-5H3v12a1 1 0 0 0 1 1h8a1 1 0 0 0 1-1M9.5 5V2.12L12.38 5zM5.13 5h-.62v1.25h2.12V5zm-.62 3h7.12v1.25H4.5zm.62 3h-.62v1.25h7.12V11z" clip-rule="evenodd" fill="#666" fill-rule="evenodd"/></svg>

demo/nextjs_voice_chat/frontend/fastrtc-demo/public/globe.svg

@@ -0,0 +1 @@
+<svg fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16"><g clip-path="url(#a)"><path fill-rule="evenodd" clip-rule="evenodd" d="M10.27 14.1a6.5 6.5 0 0 0 3.67-3.45q-1.24.21-2.7.34-.31 1.83-.97 3.1M8 16A8 8 0 1 0 8 0a8 8 0 0 0 0 16m.48-1.52a7 7 0 0 1-.96 0H7.5a4 4 0 0 1-.84-1.32q-.38-.89-.63-2.08a40 40 0 0 0 3.92 0q-.25 1.2-.63 2.08a4 4 0 0 1-.84 1.31zm2.94-4.76q1.66-.15 2.95-.43a7 7 0 0 0 0-2.58q-1.3-.27-2.95-.43a18 18 0 0 1 0 3.44m-1.27-3.54a17 17 0 0 1 0 3.64 39 39 0 0 1-4.3 0 17 17 0 0 1 0-3.64 39 39 0 0 1 4.3 0m1.1-1.17q1.45.13 2.69.34a6.5 6.5 0 0 0-3.67-3.44q.65 1.26.98 3.1M8.48 1.5l.01.02q.41.37.84 1.31.38.89.63 2.08a40 40 0 0 0-3.92 0q.25-1.2.63-2.08a4 4 0 0 1 .85-1.32 7 7 0 0 1 .96 0m-2.75.4a6.5 6.5 0 0 0-3.67 3.44 29 29 0 0 1 2.7-.34q.31-1.83.97-3.1M4.58 6.28q-1.66.16-2.95.43a7 7 0 0 0 0 2.58q1.3.27 2.95.43a18 18 0 0 1 0-3.44m.17 4.71q-1.45-.12-2.69-.34a6.5 6.5 0 0 0 3.67 3.44q-.65-1.27-.98-3.1" fill="#666"/></g><defs><clipPath id="a"><path fill="#fff" d="M0 0h16v16H0z"/></clipPath></defs></svg>

demo/nextjs_voice_chat/frontend/fastrtc-demo/public/next.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 394 80"><path fill="#000" d="M262 0h68.5v12.7h-27.2v66.6h-13.6V12.7H262V0ZM149 0v12.7H94v20.4h44.3v12.6H94v21h55v12.6H80.5V0h68.7zm34.3 0h-17.8l63.8 79.4h17.9l-32-39.7 32-39.6h-17.9l-23 28.6-23-28.6zm18.3 56.7-9-11-27.1 33.7h17.8l18.3-22.7z"/><path fill="#000" d="M81 79.3 17 0H0v79.3h13.6V17l50.2 62.3H81Zm252.6-.4c-1 0-1.8-.4-2.5-1s-1.1-1.6-1.1-2.6.3-1.8 1-2.5 1.6-1 2.6-1 1.8.3 2.5 1a3.4 3.4 0 0 1 .6 4.3 3.7 3.7 0 0 1-3 1.8zm23.2-33.5h6v23.3c0 2.1-.4 4-1.3 5.5a9.1 9.1 0 0 1-3.8 3.5c-1.6.8-3.5 1.3-5.7 1.3-2 0-3.7-.4-5.3-1s-2.8-1.8-3.7-3.2c-.9-1.3-1.4-3-1.4-5h6c.1.8.3 1.6.7 2.2s1 1.2 1.6 1.5c.7.4 1.5.5 2.4.5 1 0 1.8-.2 2.4-.6a4 4 0 0 0 1.6-1.8c.3-.8.5-1.8.5-3V45.5zm30.9 9.1a4.4 4.4 0 0 0-2-3.3 7.5 7.5 0 0 0-4.3-1.1c-1.3 0-2.4.2-3.3.5-.9.4-1.6 1-2 1.6a3.5 3.5 0 0 0-.3 4c.3.5.7.9 1.3 1.2l1.8 1 2 .5 3.2.8c1.3.3 2.5.7 3.7 1.2a13 13 0 0 1 3.2 1.8 8.1 8.1 0 0 1 3 6.5c0 2-.5 3.7-1.5 5.1a10 10 0 0 1-4.4 3.5c-1.8.8-4.1 1.2-6.8 1.2-2.6 0-4.9-.4-6.8-1.2-2-.8-3.4-2-4.5-3.5a10 10 0 0 1-1.7-5.6h6a5 5 0 0 0 3.5 4.6c1 .4 2.2.6 3.4.6 1.3 0 2.5-.2 3.5-.6 1-.4 1.8-1 2.4-1.7a4 4 0 0 0 .8-2.4c0-.9-.2-1.6-.7-2.2a11 11 0 0 0-2.1-1.4l-3.2-1-3.8-1c-2.8-.7-5-1.7-6.6-3.2a7.2 7.2 0 0 1-2.4-5.7 8 8 0 0 1 1.7-5 10 10 0 0 1 4.3-3.5c2-.8 4-1.2 6.4-1.2 2.3 0 4.4.4 6.2 1.2 1.8.8 3.2 2 4.3 3.4 1 1.4 1.5 3 1.5 5h-5.8z"/></svg>

demo/nextjs_voice_chat/frontend/fastrtc-demo/public/vercel.svg

@@ -0,0 +1 @@
+<svg fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1155 1000"><path d="m577.3 0 577.4 1000H0z" fill="#fff"/></svg>

demo/nextjs_voice_chat/frontend/fastrtc-demo/public/window.svg

@@ -0,0 +1 @@
+<svg fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16"><path fill-rule="evenodd" clip-rule="evenodd" d="M1.5 2.5h13v10a1 1 0 0 1-1 1h-11a1 1 0 0 1-1-1zM0 1h16v11.5a2.5 2.5 0 0 1-2.5 2.5h-11A2.5 2.5 0 0 1 0 12.5zm3.75 4.5a.75.75 0 1 0 0-1.5.75.75 0 0 0 0 1.5M7 4.75a.75.75 0 1 1-1.5 0 .75.75 0 0 1 1.5 0m1.75.75a.75.75 0 1 0 0-1.5.75.75 0 0 0 0 1.5" fill="#666"/></svg>

demo/nextjs_voice_chat/frontend/fastrtc-demo/tsconfig.json

@@ -0,0 +1,27 @@
+{
+  "compilerOptions": {
+    "target": "ES2017",
+    "lib": ["dom", "dom.iterable", "esnext"],
+    "allowJs": true,
+    "skipLibCheck": true,
+    "strict": true,
+    "noEmit": true,
+    "esModuleInterop": true,
+    "module": "esnext",
+    "moduleResolution": "bundler",
+    "resolveJsonModule": true,
+    "isolatedModules": true,
+    "jsx": "preserve",
+    "incremental": true,
+    "plugins": [
+      {
+        "name": "next"
+      }
+    ],
+    "paths": {
+      "@/*": ["./*"]
+    }
+  },
+  "include": ["next-env.d.ts", "**/*.ts", "**/*.tsx", ".next/types/**/*.ts"],
+  "exclude": ["node_modules"]
+}

demo/nextjs_voice_chat/requirements.txt

@@ -0,0 +1,5 @@
+openai
+fastapi
+python-dotenv
+elevenlabs
+fastrtc[vad, stt, tts]

demo/nextjs_voice_chat/run.sh

@@ -0,0 +1 @@
+uvicorn backend.server:app --host 0.0.0.0 --port 8000

demo/object_detection/README.md

@@ -0,0 +1,15 @@
+---
+title: Object Detection
+emoji: 📸
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Use YOLOv10 to detect objects in real-time
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/object_detection/app.py

@@ -0,0 +1,77 @@
+import json
+from pathlib import Path
+
+import cv2
+import gradio as gr
+from fastapi import FastAPI
+from fastapi.responses import HTMLResponse
+from fastrtc import Stream, get_twilio_turn_credentials
+from gradio.utils import get_space
+from huggingface_hub import hf_hub_download
+from pydantic import BaseModel, Field
+
+try:
+    from demo.object_detection.inference import YOLOv10
+except (ImportError, ModuleNotFoundError):
+    from inference import YOLOv10
+
+
+cur_dir = Path(__file__).parent
+
+model_file = hf_hub_download(
+    repo_id="onnx-community/yolov10n", filename="onnx/model.onnx"
+)
+
+model = YOLOv10(model_file)
+
+
+def detection(image, conf_threshold=0.3):
+    image = cv2.resize(image, (model.input_width, model.input_height))
+    print("conf_threshold", conf_threshold)
+    new_image = model.detect_objects(image, conf_threshold)
+    return cv2.resize(new_image, (500, 500))
+
+
+stream = Stream(
+    handler=detection,
+    modality="video",
+    mode="send-receive",
+    additional_inputs=[gr.Slider(minimum=0, maximum=1, step=0.01, value=0.3)],
+    rtc_configuration=get_twilio_turn_credentials() if get_space() else None,
+    concurrency_limit=2 if get_space() else None,
+)
+
+app = FastAPI()
+
+stream.mount(app)
+
+
+@app.get("/")
+async def _():
+    rtc_config = get_twilio_turn_credentials() if get_space() else None
+    html_content = open(cur_dir / "index.html").read()
+    html_content = html_content.replace("__RTC_CONFIGURATION__", json.dumps(rtc_config))
+    return HTMLResponse(content=html_content)
+
+
+class InputData(BaseModel):
+    webrtc_id: str
+    conf_threshold: float = Field(ge=0, le=1)
+
+
+@app.post("/input_hook")
+async def _(data: InputData):
+    stream.set_input(data.webrtc_id, data.conf_threshold)
+
+
+if __name__ == "__main__":
+    import os
+
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        stream.fastphone(host="0.0.0.0", port=7860)
+    else:
+        import uvicorn
+
+        uvicorn.run(app, host="0.0.0.0", port=7860)

demo/object_detection/index.html

@@ -0,0 +1,340 @@
+<!DOCTYPE html>
+<html lang="en">
+
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Object Detection</title>
+    <style>
+        body {
+            font-family: system-ui, -apple-system, sans-serif;
+            background: linear-gradient(135deg, #2d2b52 0%, #191731 100%);
+            color: white;
+            margin: 0;
+            padding: 20px;
+            height: 100vh;
+            box-sizing: border-box;
+            display: flex;
+            flex-direction: column;
+            align-items: center;
+            justify-content: center;
+        }
+
+        .container {
+            width: 100%;
+            max-width: 800px;
+            text-align: center;
+            display: flex;
+            flex-direction: column;
+            align-items: center;
+        }
+
+        .video-container {
+            width: 100%;
+            max-width: 500px;
+            aspect-ratio: 1/1;
+            background: rgba(255, 255, 255, 0.1);
+            border-radius: 12px;
+            overflow: hidden;
+            box-shadow: 0 8px 32px rgba(0, 0, 0, 0.2);
+            margin: 10px 0;
+        }
+
+        #video-output {
+            width: 100%;
+            height: 100%;
+            object-fit: cover;
+        }
+
+        button {
+            background: white;
+            color: #2d2b52;
+            border: none;
+            padding: 12px 32px;
+            border-radius: 24px;
+            font-size: 16px;
+            font-weight: 600;
+            cursor: pointer;
+            transition: all 0.3s ease;
+            box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
+        }
+
+        button:hover {
+            transform: translateY(-2px);
+            box-shadow: 0 6px 16px rgba(0, 0, 0, 0.2);
+        }
+
+        h1 {
+            font-size: 2.5em;
+            margin-bottom: 0.3em;
+        }
+
+        p {
+            color: rgba(255, 255, 255, 0.8);
+            margin-bottom: 1em;
+        }
+
+        .controls {
+            display: flex;
+            flex-direction: column;
+            gap: 12px;
+            align-items: center;
+            margin-top: 10px;
+        }
+
+        .slider-container {
+            width: 100%;
+            max-width: 300px;
+            display: flex;
+            flex-direction: column;
+            gap: 8px;
+        }
+
+        .slider-container label {
+            color: rgba(255, 255, 255, 0.8);
+            font-size: 14px;
+        }
+
+        input[type="range"] {
+            width: 100%;
+            height: 6px;
+            -webkit-appearance: none;
+            background: rgba(255, 255, 255, 0.1);
+            border-radius: 3px;
+            outline: none;
+        }
+
+        input[type="range"]::-webkit-slider-thumb {
+            -webkit-appearance: none;
+            width: 18px;
+            height: 18px;
+            background: white;
+            border-radius: 50%;
+            cursor: pointer;
+        }
+
+        /* Add styles for toast notifications */
+        .toast {
+            position: fixed;
+            top: 20px;
+            left: 50%;
+            transform: translateX(-50%);
+            padding: 16px 24px;
+            border-radius: 4px;
+            font-size: 14px;
+            z-index: 1000;
+            display: none;
+            box-shadow: 0 2px 5px rgba(0, 0, 0, 0.2);
+        }
+
+        .toast.error {
+            background-color: #f44336;
+            color: white;
+        }
+
+        .toast.warning {
+            background-color: #ffd700;
+            color: black;
+        }
+    </style>
+</head>
+
+<body>
+    <!-- Add toast element after body opening tag -->
+    <div id="error-toast" class="toast"></div>
+    <div class="container">
+        <h1>Real-time Object Detection</h1>
+        <p>Using YOLOv10 to detect objects in your webcam feed</p>
+        <div class="video-container">
+            <video id="video-output" autoplay playsinline></video>
+        </div>
+        <div class="controls">
+            <div class="slider-container">
+                <label>Confidence Threshold: <span id="conf-value">0.3</span></label>
+                <input type="range" id="conf-threshold" min="0" max="1" step="0.01" value="0.3">
+            </div>
+            <button id="start-button">Start</button>
+        </div>
+    </div>
+
+    <script>
+        let peerConnection;
+        let webrtc_id;
+        const startButton = document.getElementById('start-button');
+        const videoOutput = document.getElementById('video-output');
+        const confThreshold = document.getElementById('conf-threshold');
+        const confValue = document.getElementById('conf-value');
+
+        // Update confidence value display
+        confThreshold.addEventListener('input', (e) => {
+            confValue.textContent = e.target.value;
+            if (peerConnection) {
+                updateConfThreshold(e.target.value);
+            }
+        });
+
+        function updateConfThreshold(value) {
+            fetch('/input_hook', {
+                method: 'POST',
+                headers: {
+                    'Content-Type': 'application/json',
+                },
+                body: JSON.stringify({
+                    webrtc_id: webrtc_id,
+                    conf_threshold: parseFloat(value)
+                })
+            });
+        }
+
+        function showError(message) {
+            const toast = document.getElementById('error-toast');
+            toast.textContent = message;
+            toast.className = 'toast error';
+            toast.style.display = 'block';
+
+            // Hide toast after 5 seconds
+            setTimeout(() => {
+                toast.style.display = 'none';
+            }, 5000);
+        }
+
+        async function setupWebRTC() {
+            const config = __RTC_CONFIGURATION__;
+            peerConnection = new RTCPeerConnection(config);
+
+            const timeoutId = setTimeout(() => {
+                const toast = document.getElementById('error-toast');
+                toast.textContent = "Connection is taking longer than usual. Are you on a VPN?";
+                toast.className = 'toast warning';
+                toast.style.display = 'block';
+
+                // Hide warning after 5 seconds
+                setTimeout(() => {
+                    toast.style.display = 'none';
+                }, 5000);
+            }, 5000);
+
+            try {
+                const stream = await navigator.mediaDevices.getUserMedia({
+                    video: true
+                });
+
+                stream.getTracks().forEach(track => {
+                    peerConnection.addTrack(track, stream);
+                });
+
+                peerConnection.addEventListener('track', (evt) => {
+                    if (videoOutput && videoOutput.srcObject !== evt.streams[0]) {
+                        videoOutput.srcObject = evt.streams[0];
+                    }
+                });
+
+                const dataChannel = peerConnection.createDataChannel('text');
+                dataChannel.onmessage = (event) => {
+                    const eventJson = JSON.parse(event.data);
+                    if (eventJson.type === "error") {
+                        showError(eventJson.message);
+                    } else if (eventJson.type === "send_input") {
+                        updateConfThreshold(confThreshold.value);
+                    }
+                };
+
+                const offer = await peerConnection.createOffer();
+                await peerConnection.setLocalDescription(offer);
+
+                await new Promise((resolve) => {
+                    if (peerConnection.iceGatheringState === "complete") {
+                        resolve();
+                    } else {
+                        const checkState = () => {
+                            if (peerConnection.iceGatheringState === "complete") {
+                                peerConnection.removeEventListener("icegatheringstatechange", checkState);
+                                resolve();
+                            }
+                        };
+                        peerConnection.addEventListener("icegatheringstatechange", checkState);
+                    }
+                });
+
+                webrtc_id = Math.random().toString(36).substring(7);
+
+                const response = await fetch('/webrtc/offer', {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify({
+                        sdp: peerConnection.localDescription.sdp,
+                        type: peerConnection.localDescription.type,
+                        webrtc_id: webrtc_id
+                    })
+                });
+
+                const serverResponse = await response.json();
+
+                if (serverResponse.status === 'failed') {
+                    showError(serverResponse.meta.error === 'concurrency_limit_reached'
+                        ? `Too many connections. Maximum limit is ${serverResponse.meta.limit}`
+                        : serverResponse.meta.error);
+                    stop();
+                    startButton.textContent = 'Start';
+                    return;
+                }
+
+                await peerConnection.setRemoteDescription(serverResponse);
+
+                // Send initial confidence threshold
+                updateConfThreshold(confThreshold.value);
+
+                peerConnection.addEventListener('connectionstatechange', () => {
+                    if (peerConnection.connectionState === 'connected') {
+                        clearTimeout(timeoutId);
+                        const toast = document.getElementById('error-toast');
+                        toast.style.display = 'none';
+                    }
+                });
+
+            } catch (err) {
+                clearTimeout(timeoutId);
+                console.error('Error setting up WebRTC:', err);
+                showError('Failed to establish connection. Please try again.');
+                stop();
+                startButton.textContent = 'Start';
+            }
+        }
+
+        function stop() {
+            if (peerConnection) {
+                if (peerConnection.getTransceivers) {
+                    peerConnection.getTransceivers().forEach(transceiver => {
+                        if (transceiver.stop) {
+                            transceiver.stop();
+                        }
+                    });
+                }
+
+                if (peerConnection.getSenders) {
+                    peerConnection.getSenders().forEach(sender => {
+                        if (sender.track && sender.track.stop) sender.track.stop();
+                    });
+                }
+
+                setTimeout(() => {
+                    peerConnection.close();
+                }, 500);
+            }
+
+            videoOutput.srcObject = null;
+        }
+
+        startButton.addEventListener('click', () => {
+            if (startButton.textContent === 'Start') {
+                setupWebRTC();
+                startButton.textContent = 'Stop';
+            } else {
+                stop();
+                startButton.textContent = 'Start';
+            }
+        });
+    </script>
+</body>
+
+</html>

demo/object_detection/inference.py

@@ -0,0 +1,153 @@
+import time
+
+import cv2
+import numpy as np
+import onnxruntime
+
+try:
+    from demo.object_detection.utils import draw_detections
+except (ImportError, ModuleNotFoundError):
+    from utils import draw_detections
+
+
+class YOLOv10:
+    def __init__(self, path):
+        # Initialize model
+        self.initialize_model(path)
+
+    def __call__(self, image):
+        return self.detect_objects(image)
+
+    def initialize_model(self, path):
+        self.session = onnxruntime.InferenceSession(
+            path, providers=onnxruntime.get_available_providers()
+        )
+        # Get model info
+        self.get_input_details()
+        self.get_output_details()
+
+    def detect_objects(self, image, conf_threshold=0.3):
+        input_tensor = self.prepare_input(image)
+
+        # Perform inference on the image
+        new_image = self.inference(image, input_tensor, conf_threshold)
+
+        return new_image
+
+    def prepare_input(self, image):
+        self.img_height, self.img_width = image.shape[:2]
+
+        input_img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        # Resize input image
+        input_img = cv2.resize(input_img, (self.input_width, self.input_height))
+
+        # Scale input pixel values to 0 to 1
+        input_img = input_img / 255.0
+        input_img = input_img.transpose(2, 0, 1)
+        input_tensor = input_img[np.newaxis, :, :, :].astype(np.float32)
+
+        return input_tensor
+
+    def inference(self, image, input_tensor, conf_threshold=0.3):
+        start = time.perf_counter()
+        outputs = self.session.run(
+            self.output_names, {self.input_names[0]: input_tensor}
+        )
+
+        print(f"Inference time: {(time.perf_counter() - start) * 1000:.2f} ms")
+        (
+            boxes,
+            scores,
+            class_ids,
+        ) = self.process_output(outputs, conf_threshold)
+        return self.draw_detections(image, boxes, scores, class_ids)
+
+    def process_output(self, output, conf_threshold=0.3):
+        predictions = np.squeeze(output[0])
+
+        # Filter out object confidence scores below threshold
+        scores = predictions[:, 4]
+        predictions = predictions[scores > conf_threshold, :]
+        scores = scores[scores > conf_threshold]
+
+        if len(scores) == 0:
+            return [], [], []
+
+        # Get the class with the highest confidence
+        class_ids = predictions[:, 5].astype(int)
+
+        # Get bounding boxes for each object
+        boxes = self.extract_boxes(predictions)
+
+        return boxes, scores, class_ids
+
+    def extract_boxes(self, predictions):
+        # Extract boxes from predictions
+        boxes = predictions[:, :4]
+
+        # Scale boxes to original image dimensions
+        boxes = self.rescale_boxes(boxes)
+
+        # Convert boxes to xyxy format
+        # boxes = xywh2xyxy(boxes)
+
+        return boxes
+
+    def rescale_boxes(self, boxes):
+        # Rescale boxes to original image dimensions
+        input_shape = np.array(
+            [self.input_width, self.input_height, self.input_width, self.input_height]
+        )
+        boxes = np.divide(boxes, input_shape, dtype=np.float32)
+        boxes *= np.array(
+            [self.img_width, self.img_height, self.img_width, self.img_height]
+        )
+        return boxes
+
+    def draw_detections(
+        self, image, boxes, scores, class_ids, draw_scores=True, mask_alpha=0.4
+    ):
+        return draw_detections(image, boxes, scores, class_ids, mask_alpha)
+
+    def get_input_details(self):
+        model_inputs = self.session.get_inputs()
+        self.input_names = [model_inputs[i].name for i in range(len(model_inputs))]
+
+        self.input_shape = model_inputs[0].shape
+        self.input_height = self.input_shape[2]
+        self.input_width = self.input_shape[3]
+
+    def get_output_details(self):
+        model_outputs = self.session.get_outputs()
+        self.output_names = [model_outputs[i].name for i in range(len(model_outputs))]
+
+
+if __name__ == "__main__":
+    import tempfile
+
+    import requests
+    from huggingface_hub import hf_hub_download
+
+    model_file = hf_hub_download(
+        repo_id="onnx-community/yolov10s", filename="onnx/model.onnx"
+    )
+
+    yolov8_detector = YOLOv10(model_file)
+
+    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
+        f.write(
+            requests.get(
+                "https://live.staticflickr.com/13/19041780_d6fd803de0_3k.jpg"
+            ).content
+        )
+        f.seek(0)
+        img = cv2.imread(f.name)
+
+    # # Detect Objects
+    combined_image = yolov8_detector.detect_objects(img)
+
+    # Draw detections
+    cv2.namedWindow("Output", cv2.WINDOW_NORMAL)
+    cv2.imshow("Output", combined_image)
+    cv2.waitKey(0)

demo/object_detection/requirements.txt

@@ -0,0 +1,4 @@
+fastrtc
+opencv-python
+twilio
+onnxruntime-gpu

demo/object_detection/utils.py

@@ -0,0 +1,237 @@
+import cv2
+import numpy as np
+
+class_names = [
+    "person",
+    "bicycle",
+    "car",
+    "motorcycle",
+    "airplane",
+    "bus",
+    "train",
+    "truck",
+    "boat",
+    "traffic light",
+    "fire hydrant",
+    "stop sign",
+    "parking meter",
+    "bench",
+    "bird",
+    "cat",
+    "dog",
+    "horse",
+    "sheep",
+    "cow",
+    "elephant",
+    "bear",
+    "zebra",
+    "giraffe",
+    "backpack",
+    "umbrella",
+    "handbag",
+    "tie",
+    "suitcase",
+    "frisbee",
+    "skis",
+    "snowboard",
+    "sports ball",
+    "kite",
+    "baseball bat",
+    "baseball glove",
+    "skateboard",
+    "surfboard",
+    "tennis racket",
+    "bottle",
+    "wine glass",
+    "cup",
+    "fork",
+    "knife",
+    "spoon",
+    "bowl",
+    "banana",
+    "apple",
+    "sandwich",
+    "orange",
+    "broccoli",
+    "carrot",
+    "hot dog",
+    "pizza",
+    "donut",
+    "cake",
+    "chair",
+    "couch",
+    "potted plant",
+    "bed",
+    "dining table",
+    "toilet",
+    "tv",
+    "laptop",
+    "mouse",
+    "remote",
+    "keyboard",
+    "cell phone",
+    "microwave",
+    "oven",
+    "toaster",
+    "sink",
+    "refrigerator",
+    "book",
+    "clock",
+    "vase",
+    "scissors",
+    "teddy bear",
+    "hair drier",
+    "toothbrush",
+]
+
+# Create a list of colors for each class where each color is a tuple of 3 integer values
+rng = np.random.default_rng(3)
+colors = rng.uniform(0, 255, size=(len(class_names), 3))
+
+
+def nms(boxes, scores, iou_threshold):
+    # Sort by score
+    sorted_indices = np.argsort(scores)[::-1]
+
+    keep_boxes = []
+    while sorted_indices.size > 0:
+        # Pick the last box
+        box_id = sorted_indices[0]
+        keep_boxes.append(box_id)
+
+        # Compute IoU of the picked box with the rest
+        ious = compute_iou(boxes[box_id, :], boxes[sorted_indices[1:], :])
+
+        # Remove boxes with IoU over the threshold
+        keep_indices = np.where(ious < iou_threshold)[0]
+
+        # print(keep_indices.shape, sorted_indices.shape)
+        sorted_indices = sorted_indices[keep_indices + 1]
+
+    return keep_boxes
+
+
+def multiclass_nms(boxes, scores, class_ids, iou_threshold):
+    unique_class_ids = np.unique(class_ids)
+
+    keep_boxes = []
+    for class_id in unique_class_ids:
+        class_indices = np.where(class_ids == class_id)[0]
+        class_boxes = boxes[class_indices, :]
+        class_scores = scores[class_indices]
+
+        class_keep_boxes = nms(class_boxes, class_scores, iou_threshold)
+        keep_boxes.extend(class_indices[class_keep_boxes])
+
+    return keep_boxes
+
+
+def compute_iou(box, boxes):
+    # Compute xmin, ymin, xmax, ymax for both boxes
+    xmin = np.maximum(box[0], boxes[:, 0])
+    ymin = np.maximum(box[1], boxes[:, 1])
+    xmax = np.minimum(box[2], boxes[:, 2])
+    ymax = np.minimum(box[3], boxes[:, 3])
+
+    # Compute intersection area
+    intersection_area = np.maximum(0, xmax - xmin) * np.maximum(0, ymax - ymin)
+
+    # Compute union area
+    box_area = (box[2] - box[0]) * (box[3] - box[1])
+    boxes_area = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
+    union_area = box_area + boxes_area - intersection_area
+
+    # Compute IoU
+    iou = intersection_area / union_area
+
+    return iou
+
+
+def xywh2xyxy(x):
+    # Convert bounding box (x, y, w, h) to bounding box (x1, y1, x2, y2)
+    y = np.copy(x)
+    y[..., 0] = x[..., 0] - x[..., 2] / 2
+    y[..., 1] = x[..., 1] - x[..., 3] / 2
+    y[..., 2] = x[..., 0] + x[..., 2] / 2
+    y[..., 3] = x[..., 1] + x[..., 3] / 2
+    return y
+
+
+def draw_detections(image, boxes, scores, class_ids, mask_alpha=0.3):
+    det_img = image.copy()
+
+    img_height, img_width = image.shape[:2]
+    font_size = min([img_height, img_width]) * 0.0006
+    text_thickness = int(min([img_height, img_width]) * 0.001)
+
+    # det_img = draw_masks(det_img, boxes, class_ids, mask_alpha)
+
+    # Draw bounding boxes and labels of detections
+    for class_id, box, score in zip(class_ids, boxes, scores):
+        color = colors[class_id]
+
+        draw_box(det_img, box, color)  # type: ignore
+
+        label = class_names[class_id]
+        caption = f"{label} {int(score * 100)}%"
+        draw_text(det_img, caption, box, color, font_size, text_thickness)  # type: ignore
+
+    return det_img
+
+
+def draw_box(
+    image: np.ndarray,
+    box: np.ndarray,
+    color: tuple[int, int, int] = (0, 0, 255),
+    thickness: int = 2,
+) -> np.ndarray:
+    x1, y1, x2, y2 = box.astype(int)
+    return cv2.rectangle(image, (x1, y1), (x2, y2), color, thickness)
+
+
+def draw_text(
+    image: np.ndarray,
+    text: str,
+    box: np.ndarray,
+    color: tuple[int, int, int] = (0, 0, 255),
+    font_size: float = 0.001,
+    text_thickness: int = 2,
+) -> np.ndarray:
+    x1, y1, x2, y2 = box.astype(int)
+    (tw, th), _ = cv2.getTextSize(
+        text=text,
+        fontFace=cv2.FONT_HERSHEY_SIMPLEX,
+        fontScale=font_size,
+        thickness=text_thickness,
+    )
+    th = int(th * 1.2)
+
+    cv2.rectangle(image, (x1, y1), (x1 + tw, y1 - th), color, -1)
+
+    return cv2.putText(
+        image,
+        text,
+        (x1, y1),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        font_size,
+        (255, 255, 255),
+        text_thickness,
+        cv2.LINE_AA,
+    )
+
+
+def draw_masks(
+    image: np.ndarray, boxes: np.ndarray, classes: np.ndarray, mask_alpha: float = 0.3
+) -> np.ndarray:
+    mask_img = image.copy()
+
+    # Draw bounding boxes and labels of detections
+    for box, class_id in zip(boxes, classes):
+        color = colors[class_id]
+
+        x1, y1, x2, y2 = box.astype(int)
+
+        # Draw fill rectangle in mask image
+        cv2.rectangle(mask_img, (x1, y1), (x2, y2), color, -1)  # type: ignore
+
+    return cv2.addWeighted(mask_img, mask_alpha, image, 1 - mask_alpha, 0)

demo/phonic_chat/README.md

@@ -0,0 +1,16 @@
+---
+title: Phonic AI Chat
+emoji: 🎙️
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Talk to Phonic AI's speech-to-speech model
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|PHONIC_API_KEY]
+python_version: 3.11
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/phonic_chat/app.py

@@ -0,0 +1,116 @@
+import asyncio
+import base64
+import os
+
+import gradio as gr
+from gradio.utils import get_space
+import numpy as np
+from dotenv import load_dotenv
+from fastrtc import (
+    AdditionalOutputs,
+    AsyncStreamHandler,
+    Stream,
+    get_twilio_turn_credentials,
+    audio_to_float32,
+    wait_for_item,
+)
+from phonic.client import PhonicSTSClient, get_voices
+
+load_dotenv()
+
+STS_URI = "wss://api.phonic.co/v1/sts/ws"
+API_KEY = os.environ["PHONIC_API_KEY"]
+SAMPLE_RATE = 44_100
+voices = get_voices(API_KEY)
+voice_ids = [voice["id"] for voice in voices]
+
+
+class PhonicHandler(AsyncStreamHandler):
+    def __init__(self):
+        super().__init__(input_sample_rate=SAMPLE_RATE, output_sample_rate=SAMPLE_RATE)
+        self.output_queue = asyncio.Queue()
+        self.client = None
+
+    def copy(self) -> AsyncStreamHandler:
+        return PhonicHandler()
+
+    async def start_up(self):
+        await self.wait_for_args()
+        voice_id = self.latest_args[1]
+        async with PhonicSTSClient(STS_URI, API_KEY) as client:
+            self.client = client
+            sts_stream = client.sts(  # type: ignore
+                input_format="pcm_44100",
+                output_format="pcm_44100",
+                system_prompt="You are a helpful voice assistant. Respond conversationally.",
+                # welcome_message="Hello! I'm your voice assistant. How can I help you today?",
+                voice_id=voice_id,
+            )
+            async for message in sts_stream:
+                message_type = message.get("type")
+                if message_type == "audio_chunk":
+                    audio_b64 = message["audio"]
+                    audio_bytes = base64.b64decode(audio_b64)
+                    await self.output_queue.put(
+                        (SAMPLE_RATE, np.frombuffer(audio_bytes, dtype=np.int16))
+                    )
+                    if text := message.get("text"):
+                        msg = {"role": "assistant", "content": text}
+                        await self.output_queue.put(AdditionalOutputs(msg))
+                elif message_type == "input_text":
+                    msg = {"role": "user", "content": message["text"]}
+                    await self.output_queue.put(AdditionalOutputs(msg))
+
+    async def emit(self):
+        return await wait_for_item(self.output_queue)
+
+    async def receive(self, frame: tuple[int, np.ndarray]) -> None:
+        if not self.client:
+            return
+        audio_float32 = audio_to_float32(frame)
+        await self.client.send_audio(audio_float32)  # type: ignore
+
+    async def shutdown(self):
+        if self.client:
+            await self.client._websocket.close()
+        return super().shutdown()
+
+
+def add_to_chatbot(chatbot, message):
+    chatbot.append(message)
+    return chatbot
+
+
+chatbot = gr.Chatbot(type="messages", value=[])
+stream = Stream(
+    handler=PhonicHandler(),
+    mode="send-receive",
+    modality="audio",
+    additional_inputs=[
+        gr.Dropdown(
+            choices=voice_ids,
+            value="victoria",
+            label="Voice",
+            info="Select a voice from the dropdown",
+        )
+    ],
+    additional_outputs=[chatbot],
+    additional_outputs_handler=add_to_chatbot,
+    ui_args={
+        "title": "Phonic Chat (Powered by FastRTC ⚡️)",
+    },
+    rtc_configuration=get_twilio_turn_credentials() if get_space() else None,
+    concurrency_limit=5 if get_space() else None,
+    time_limit=90 if get_space() else None,
+)
+
+# with stream.ui:
+#     state.change(lambda s: s, inputs=state, outputs=chatbot)
+
+if __name__ == "__main__":
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        stream.fastphone(host="0.0.0.0", port=7860)
+    else:
+        stream.ui.launch(server_port=7860)

demo/phonic_chat/requirements.txt

@@ -0,0 +1,74 @@
+# This file was autogenerated by uv via the following command:
+#    uv pip compile requirements.in -o requirements.txt
+aiohappyeyeballs==2.4.6
+    # via aiohttp
+aiohttp==3.11.12
+    # via
+    #   aiohttp-retry
+    #   twilio
+aiohttp-retry==2.9.1
+    # via twilio
+aiosignal==1.3.2
+    # via aiohttp
+attrs==25.1.0
+    # via aiohttp
+certifi==2025.1.31
+    # via requests
+cffi==1.17.1
+    # via sounddevice
+charset-normalizer==3.4.1
+    # via requests
+fastrtc==0.0.1
+    # via -r requirements.in
+frozenlist==1.5.0
+    # via
+    #   aiohttp
+    #   aiosignal
+idna==3.10
+    # via
+    #   requests
+    #   yarl
+isort==6.0.0
+    # via phonic-python
+loguru==0.7.3
+    # via phonic-python
+multidict==6.1.0
+    # via
+    #   aiohttp
+    #   yarl
+numpy==2.2.3
+    # via
+    #   phonic-python
+    #   scipy
+phonic-python==0.1.3
+    # via -r requirements.in
+propcache==0.3.0
+    # via
+    #   aiohttp
+    #   yarl
+pycparser==2.22
+    # via cffi
+pyjwt==2.10.1
+    # via twilio
+python-dotenv==1.0.1
+    # via
+    #   -r requirements.in
+    #   phonic-python
+requests==2.32.3
+    # via
+    #   phonic-python
+    #   twilio
+scipy==1.15.2
+    # via phonic-python
+sounddevice==0.5.1
+    # via phonic-python
+twilio==9.4.6
+    # via -r requirements.in
+typing-extensions==4.12.2
+    # via phonic-python
+urllib3==2.3.0
+    # via requests
+websockets==15.0
+    # via phonic-python
+yarl==1.18.3
+    # via aiohttp

demo/talk_to_claude/README.md

@@ -0,0 +1,15 @@
+---
+title: Talk to Claude
+emoji: 👨‍🦰
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Talk to Anthropic's Claude
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GROQ_API_KEY, secret|ANTHROPIC_API_KEY, secret|ELEVENLABS_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/talk_to_claude/app.py

@@ -0,0 +1,134 @@
+import json
+import os
+import time
+from pathlib import Path
+
+import anthropic
+import gradio as gr
+import numpy as np
+from dotenv import load_dotenv
+from elevenlabs import ElevenLabs
+from fastapi import FastAPI
+from fastapi.responses import HTMLResponse, StreamingResponse
+from fastrtc import (
+    AdditionalOutputs,
+    ReplyOnPause,
+    Stream,
+    get_tts_model,
+    get_twilio_turn_credentials,
+)
+from fastrtc.utils import audio_to_bytes
+from gradio.utils import get_space
+from groq import Groq
+from pydantic import BaseModel
+
+load_dotenv()
+
+groq_client = Groq()
+claude_client = anthropic.Anthropic()
+tts_client = ElevenLabs(api_key=os.environ["ELEVENLABS_API_KEY"])
+
+curr_dir = Path(__file__).parent
+
+tts_model = get_tts_model()
+
+
+def response(
+    audio: tuple[int, np.ndarray],
+    chatbot: list[dict] | None = None,
+):
+    chatbot = chatbot or []
+    messages = [{"role": d["role"], "content": d["content"]} for d in chatbot]
+    prompt = groq_client.audio.transcriptions.create(
+        file=("audio-file.mp3", audio_to_bytes(audio)),
+        model="whisper-large-v3-turbo",
+        response_format="verbose_json",
+    ).text
+    chatbot.append({"role": "user", "content": prompt})
+    yield AdditionalOutputs(chatbot)
+    messages.append({"role": "user", "content": prompt})
+    response = claude_client.messages.create(
+        model="claude-3-5-haiku-20241022",
+        max_tokens=512,
+        messages=messages,  # type: ignore
+    )
+    response_text = " ".join(
+        block.text  # type: ignore
+        for block in response.content
+        if getattr(block, "type", None) == "text"
+    )
+    chatbot.append({"role": "assistant", "content": response_text})
+
+    start = time.time()
+
+    print("starting tts", start)
+    for i, chunk in enumerate(tts_model.stream_tts_sync(response_text)):
+        print("chunk", i, time.time() - start)
+        yield chunk
+        print("finished tts", time.time() - start)
+        yield AdditionalOutputs(chatbot)
+
+
+chatbot = gr.Chatbot(type="messages")
+stream = Stream(
+    modality="audio",
+    mode="send-receive",
+    handler=ReplyOnPause(response),
+    additional_outputs_handler=lambda a, b: b,
+    additional_inputs=[chatbot],
+    additional_outputs=[chatbot],
+    rtc_configuration=get_twilio_turn_credentials() if get_space() else None,
+    concurrency_limit=5 if get_space() else None,
+    time_limit=90 if get_space() else None,
+)
+
+
+class Message(BaseModel):
+    role: str
+    content: str
+
+
+class InputData(BaseModel):
+    webrtc_id: str
+    chatbot: list[Message]
+
+
+app = FastAPI()
+stream.mount(app)
+
+
+@app.get("/")
+async def _():
+    rtc_config = get_twilio_turn_credentials() if get_space() else None
+    html_content = (curr_dir / "index.html").read_text()
+    html_content = html_content.replace("__RTC_CONFIGURATION__", json.dumps(rtc_config))
+    return HTMLResponse(content=html_content, status_code=200)
+
+
+@app.post("/input_hook")
+async def _(body: InputData):
+    stream.set_input(body.webrtc_id, body.model_dump()["chatbot"])
+    return {"status": "ok"}
+
+
+@app.get("/outputs")
+def _(webrtc_id: str):
+    async def output_stream():
+        async for output in stream.output_stream(webrtc_id):
+            chatbot = output.args[0]
+            yield f"event: output\ndata: {json.dumps(chatbot[-1])}\n\n"
+
+    return StreamingResponse(output_stream(), media_type="text/event-stream")
+
+
+if __name__ == "__main__":
+    import os
+
+    if (mode := os.getenv("MODE")) == "UI":
+        stream.ui.launch(server_port=7860)
+    elif mode == "PHONE":
+        stream.fastphone(host="0.0.0.0", port=7860)
+    else:
+        import uvicorn
+
+        uvicorn.run(app, host="0.0.0.0", port=7860)

demo/talk_to_claude/index.html

@@ -0,0 +1,546 @@
+<!DOCTYPE html>
+<html lang="en">
+
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>RetroChat Audio</title>
+    <style>
+        body {
+            font-family: monospace;
+            background-color: #1a1a1a;
+            color: #00ff00;
+            margin: 0;
+            padding: 20px;
+            height: 100vh;
+            box-sizing: border-box;
+        }
+
+        .container {
+            display: flex;
+            flex-direction: column;
+            gap: 20px;
+            height: calc(100% - 100px);
+            margin-bottom: 20px;
+        }
+
+        .chat-container {
+            border: 2px solid #00ff00;
+            padding: 20px;
+            display: flex;
+            flex-direction: column;
+            flex-grow: 1;
+            box-sizing: border-box;
+        }
+
+        .controls-container {
+            border: 2px solid #00ff00;
+            padding: 20px;
+            display: flex;
+            align-items: center;
+            gap: 20px;
+            height: 128px;
+            box-sizing: border-box;
+        }
+
+        .visualization-container {
+            flex-grow: 1;
+            display: flex;
+            align-items: center;
+        }
+
+        .box-container {
+            display: flex;
+            justify-content: space-between;
+            height: 64px;
+            width: 100%;
+        }
+
+        .box {
+            height: 100%;
+            width: 8px;
+            background: #00ff00;
+            border-radius: 8px;
+            transition: transform 0.05s ease;
+        }
+
+        .chat-messages {
+            flex-grow: 1;
+            overflow-y: auto;
+            margin-bottom: 20px;
+            padding: 10px;
+            border: 1px solid #00ff00;
+        }
+
+        .message {
+            margin-bottom: 10px;
+            padding: 8px;
+            border-radius: 4px;
+        }
+
+        .message.user {
+            background-color: #003300;
+        }
+
+        .message.assistant {
+            background-color: #002200;
+        }
+
+        button {
+            height: 64px;
+            min-width: 120px;
+            background-color: #000;
+            color: #00ff00;
+            border: 2px solid #00ff00;
+            padding: 10px 20px;
+            font-family: monospace;
+            font-size: 16px;
+            cursor: pointer;
+            transition: all 0.3s;
+        }
+
+        button:hover {
+            border-width: 3px;
+        }
+
+        #audio-output {
+            display: none;
+        }
+
+        /* Retro CRT effect */
+        .crt-overlay {
+            position: absolute;
+            top: 0;
+            left: 0;
+            width: 100%;
+            height: 100%;
+            background: repeating-linear-gradient(0deg,
+                    rgba(0, 255, 0, 0.03),
+                    rgba(0, 255, 0, 0.03) 1px,
+                    transparent 1px,
+                    transparent 2px);
+            pointer-events: none;
+        }
+
+        /* Add these new styles */
+        .icon-with-spinner {
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            gap: 12px;
+            min-width: 180px;
+        }
+
+        .spinner {
+            width: 20px;
+            height: 20px;
+            border: 2px solid #00ff00;
+            border-top-color: transparent;
+            border-radius: 50%;
+            animation: spin 1s linear infinite;
+            flex-shrink: 0;
+        }
+
+        @keyframes spin {
+            to {
+                transform: rotate(360deg);
+            }
+        }
+
+        .pulse-container {
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            gap: 12px;
+            min-width: 180px;
+        }
+
+        .pulse-circle {
+            width: 20px;
+            height: 20px;
+            border-radius: 50%;
+            background-color: #00ff00;
+            opacity: 0.2;
+            flex-shrink: 0;
+            transform: translateX(-0%) scale(var(--audio-level, 1));
+            transition: transform 0.1s ease;
+        }
+
+        /* Add styles for typing indicator */
+        .typing-indicator {
+            padding: 8px;
+            background-color: #002200;
+            border-radius: 4px;
+            margin-bottom: 10px;
+            display: none;
+        }
+
+        .dots {
+            display: inline-flex;
+            gap: 4px;
+        }
+
+        .dot {
+            width: 8px;
+            height: 8px;
+            background-color: #00ff00;
+            border-radius: 50%;
+            animation: pulse 1.5s infinite;
+            opacity: 0.5;
+        }
+
+        .dot:nth-child(2) {
+            animation-delay: 0.5s;
+        }
+
+        .dot:nth-child(3) {
+            animation-delay: 1s;
+        }
+
+        @keyframes pulse {
+
+            0%,
+            100% {
+                opacity: 0.5;
+                transform: scale(1);
+            }
+
+            50% {
+                opacity: 1;
+                transform: scale(1.2);
+            }
+        }
+
+        /* Add styles for toast notifications */
+        .toast {
+            position: fixed;
+            top: 20px;
+            left: 50%;
+            transform: translateX(-50%);
+            padding: 16px 24px;
+            border-radius: 4px;
+            font-size: 14px;
+            z-index: 1000;
+            display: none;
+            box-shadow: 0 2px 5px rgba(0, 0, 0, 0.2);
+        }
+
+        .toast.error {
+            background-color: #f44336;
+            color: white;
+        }
+
+        .toast.warning {
+            background-color: #ffd700;
+            color: black;
+        }
+    </style>
+</head>
+
+<body>
+    <!-- Add toast element after body opening tag -->
+    <div id="error-toast" class="toast"></div>
+    <div class="container">
+        <div class="chat-container">
+            <div class="chat-messages" id="chat-messages"></div>
+            <!-- Move typing indicator outside the chat messages -->
+            <div class="typing-indicator" id="typing-indicator">
+                <div class="dots">
+                    <div class="dot"></div>
+                    <div class="dot"></div>
+                    <div class="dot"></div>
+                </div>
+            </div>
+        </div>
+        <div class="controls-container">
+            <div class="visualization-container">
+                <div class="box-container">
+                    <!-- Boxes will be dynamically added here -->
+                </div>
+            </div>
+            <button id="start-button">Start</button>
+        </div>
+    </div>
+    <audio id="audio-output"></audio>
+
+    <script>
+        let audioContext;
+        let analyser_input, analyser_output;
+        let dataArray_input, dataArray_output;
+        let animationId_input, animationId_output;
+        let chatHistory = [];
+        let peerConnection;
+        let webrtc_id;
+
+        const audioOutput = document.getElementById('audio-output');
+        const startButton = document.getElementById('start-button');
+        const chatMessages = document.getElementById('chat-messages');
+
+        function updateButtonState() {
+            if (peerConnection && (peerConnection.connectionState === 'connecting' || peerConnection.connectionState === 'new')) {
+                startButton.innerHTML = `
+                    <div class="icon-with-spinner">
+                        <div class="spinner"></div>
+                        <span>Connecting...</span>
+                    </div>
+                `;
+            } else if (peerConnection && peerConnection.connectionState === 'connected') {
+                startButton.innerHTML = `
+                    <div class="pulse-container">
+                        <div class="pulse-circle"></div>
+                        <span>Stop</span>
+                    </div>
+                `;
+            } else {
+                startButton.innerHTML = 'Start';
+            }
+        }
+
+        function showError(message) {
+            const toast = document.getElementById('error-toast');
+            toast.textContent = message;
+            toast.className = 'toast error';
+            toast.style.display = 'block';
+
+            // Hide toast after 5 seconds
+            setTimeout(() => {
+                toast.style.display = 'none';
+            }, 5000);
+        }
+
+        async function setupWebRTC() {
+            const config = __RTC_CONFIGURATION__;
+            peerConnection = new RTCPeerConnection(config);
+
+            const timeoutId = setTimeout(() => {
+                const toast = document.getElementById('error-toast');
+                toast.textContent = "Connection is taking longer than usual. Are you on a VPN?";
+                toast.className = 'toast warning';
+                toast.style.display = 'block';
+
+                // Hide warning after 5 seconds
+                setTimeout(() => {
+                    toast.style.display = 'none';
+                }, 5000);
+            }, 5000);
+
+            try {
+                const stream = await navigator.mediaDevices.getUserMedia({
+                    audio: true
+                });
+
+                // Set up input visualization
+                audioContext = new AudioContext();
+                analyser_input = audioContext.createAnalyser();
+                const inputSource = audioContext.createMediaStreamSource(stream);
+                inputSource.connect(analyser_input);
+                analyser_input.fftSize = 64;
+                dataArray_input = new Uint8Array(analyser_input.frequencyBinCount);
+
+                function updateAudioLevel() {
+                    analyser_input.getByteFrequencyData(dataArray_input);
+                    const average = Array.from(dataArray_input).reduce((a, b) => a + b, 0) / dataArray_input.length;
+                    const audioLevel = average / 255;
+
+                    const pulseCircle = document.querySelector('.pulse-circle');
+                    if (pulseCircle) {
+                        pulseCircle.style.setProperty('--audio-level', 1 + audioLevel);
+                    }
+
+                    animationId_input = requestAnimationFrame(updateAudioLevel);
+                }
+                updateAudioLevel();
+
+                stream.getTracks().forEach(track => {
+                    peerConnection.addTrack(track, stream);
+                });
+
+                // Add connection state change listener
+                peerConnection.addEventListener('connectionstatechange', () => {
+                    console.log('Connection state:', peerConnection.connectionState);
+                    if (peerConnection.connectionState === 'connected') {
+                        clearTimeout(timeoutId);
+                        const toast = document.getElementById('error-toast');
+                        toast.style.display = 'none';
+                    }
+                    updateButtonState();
+                });
+
+                // Handle incoming audio
+                peerConnection.addEventListener('track', (evt) => {
+                    if (audioOutput.srcObject !== evt.streams[0]) {
+                        audioOutput.srcObject = evt.streams[0];
+                        audioOutput.play();
+
+                        // Set up output visualization
+                        analyser_output = audioContext.createAnalyser();
+                        const outputSource = audioContext.createMediaStreamSource(evt.streams[0]);
+                        outputSource.connect(analyser_output);
+                        analyser_output.fftSize = 2048;
+                        dataArray_output = new Uint8Array(analyser_output.frequencyBinCount);
+                        updateVisualization();
+                    }
+                });
+
+                // Create data channel for messages
+                const dataChannel = peerConnection.createDataChannel('text');
+                dataChannel.onmessage = (event) => {
+                    const eventJson = JSON.parse(event.data);
+                    const typingIndicator = document.getElementById('typing-indicator');
+
+                    if (eventJson.type === "error") {
+                        showError(eventJson.message);
+                    } else if (eventJson.type === "send_input") {
+                        fetch('/input_hook', {
+                            method: 'POST',
+                            headers: {
+                                'Content-Type': 'application/json',
+                            },
+                            body: JSON.stringify({
+                                webrtc_id: webrtc_id,
+                                chatbot: chatHistory
+                            })
+                        });
+                    } else if (eventJson.type === "log") {
+                        if (eventJson.data === "pause_detected") {
+                            typingIndicator.style.display = 'block';
+                            chatMessages.scrollTop = chatMessages.scrollHeight;
+                        } else if (eventJson.data === "response_starting") {
+                            typingIndicator.style.display = 'none';
+                        }
+                    }
+                };
+
+                // Create and send offer
+                const offer = await peerConnection.createOffer();
+                await peerConnection.setLocalDescription(offer);
+
+                await new Promise((resolve) => {
+                    if (peerConnection.iceGatheringState === "complete") {
+                        resolve();
+                    } else {
+                        const checkState = () => {
+                            if (peerConnection.iceGatheringState === "complete") {
+                                peerConnection.removeEventListener("icegatheringstatechange", checkState);
+                                resolve();
+                            }
+                        };
+                        peerConnection.addEventListener("icegatheringstatechange", checkState);
+                    }
+                });
+
+                webrtc_id = Math.random().toString(36).substring(7);
+
+                const response = await fetch('/webrtc/offer', {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify({
+                        sdp: peerConnection.localDescription.sdp,
+                        type: peerConnection.localDescription.type,
+                        webrtc_id: webrtc_id
+                    })
+                });
+
+                const serverResponse = await response.json();
+
+                if (serverResponse.status === 'failed') {
+                    showError(serverResponse.meta.error === 'concurrency_limit_reached'
+                        ? `Too many connections. Maximum limit is ${serverResponse.meta.limit}`
+                        : serverResponse.meta.error);
+                    stop();
+                    return;
+                }
+
+                await peerConnection.setRemoteDescription(serverResponse);
+
+                // Start visualization
+                updateVisualization();
+
+                // create event stream to receive messages from /output
+                const eventSource = new EventSource('/outputs?webrtc_id=' + webrtc_id);
+                eventSource.addEventListener("output", (event) => {
+                    const eventJson = JSON.parse(event.data);
+                    addMessage(eventJson.role, eventJson.content);
+                });
+            } catch (err) {
+                clearTimeout(timeoutId);
+                console.error('Error setting up WebRTC:', err);
+                showError('Failed to establish connection. Please try again.');
+                stop();
+            }
+        }
+
+        function addMessage(role, content) {
+            const messageDiv = document.createElement('div');
+            messageDiv.classList.add('message', role);
+            messageDiv.textContent = content;
+            chatMessages.appendChild(messageDiv);
+            chatMessages.scrollTop = chatMessages.scrollHeight;
+            chatHistory.push({ role, content });
+        }
+
+        // Add this after other const declarations
+        const boxContainer = document.querySelector('.box-container');
+        const numBars = 32;
+        for (let i = 0; i < numBars; i++) {
+            const box = document.createElement('div');
+            box.className = 'box';
+            boxContainer.appendChild(box);
+        }
+
+        // Replace the draw function with updateVisualization
+        function updateVisualization() {
+            animationId_output = requestAnimationFrame(updateVisualization);
+
+            analyser_output.getByteFrequencyData(dataArray_output);
+            const bars = document.querySelectorAll('.box');
+
+            for (let i = 0; i < bars.length; i++) {
+                const barHeight = (dataArray_output[i] / 255) * 2;
+                bars[i].style.transform = `scaleY(${Math.max(0.1, barHeight)})`;
+            }
+        }
+
+        function stop() {
+            if (peerConnection) {
+                if (peerConnection.getTransceivers) {
+                    peerConnection.getTransceivers().forEach(transceiver => {
+                        if (transceiver.stop) {
+                            transceiver.stop();
+                        }
+                    });
+                }
+
+                if (peerConnection.getSenders) {
+                    peerConnection.getSenders().forEach(sender => {
+                        if (sender.track && sender.track.stop) sender.track.stop();
+                    });
+                }
+
+                peerConnection.close();
+            }
+
+            if (animationId_input) {
+                cancelAnimationFrame(animationId_input);
+            }
+            if (animationId_output) {
+                cancelAnimationFrame(animationId_output);
+            }
+            if (audioContext) {
+                audioContext.close();
+            }
+
+            updateButtonState();
+        }
+
+        startButton.addEventListener('click', () => {
+            if (startButton.textContent === 'Start') {
+                setupWebRTC();
+            } else {
+                stop();
+            }
+        });
+    </script>
+</body>
+
+</html>

demo/talk_to_claude/requirements.txt

@@ -0,0 +1,6 @@
+fastrtc[vad, tts]
+elevenlabs
+groq
+anthropic
+twilio
+python-dotenv

demo/talk_to_gemini/README.md

@@ -0,0 +1,15 @@
+---
+title: Talk to Gemini
+emoji: ♊️
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Talk to Gemini using Google's multimodal API
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GEMINI_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

demo/talk_to_gemini/README_gradio.md

@@ -0,0 +1,15 @@
+---
+title: Talk to Gemini (Gradio UI)
+emoji: ♊️
+colorFrom: purple
+colorTo: red
+sdk: gradio
+sdk_version: 5.16.0
+app_file: app.py
+pinned: false
+license: mit
+short_description: Talk to Gemini (Gradio UI)
+tags: [webrtc, websocket, gradio, secret|TWILIO_ACCOUNT_SID, secret|TWILIO_AUTH_TOKEN, secret|GEMINI_API_KEY]
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference