Merge branch 'master' of https://github.com/snakers4/silero-vad into max_speech_duration_v4

README.md

@@ -15,7 +15,7 @@ This repository also includes Number Detector and Language classifier [models](h
 <br/>
 
 <p align="center">
-  <img src="https://user-images.githubusercontent.com/36505480/145563071-681b57e3-06b5-4cd0-bdee-e2ade3d50a60.png" />
+  <img src="https://user-images.githubusercontent.com/36505480/198026365-8da383e0-5398-4a12-b7f8-22c2c0059512.png" />
 </p>
 
 <details>
@@ -35,11 +35,11 @@ https://user-images.githubusercontent.com/36505480/144874384-95f80f6d-a4f1-42cc-
   
 - **Fast**
 
-  One audio chunk (30+ ms) [takes](https://github.com/snakers4/silero-vad/wiki/Performance-Metrics#silero-vad-performance-metrics) around **1ms** to be processed on a single CPU thread. Using batching or GPU can also improve performance considerably. Under certain conditions ONNX may even run up to 2-3x faster. 
+  One audio chunk (30+ ms) [takes](https://github.com/snakers4/silero-vad/wiki/Performance-Metrics#silero-vad-performance-metrics) less than **1ms** to be processed on a single CPU thread. Using batching or GPU can also improve performance considerably. Under certain conditions ONNX may even run up to 4-5x faster. 
 
 - **Lightweight**
 
-  JIT model is less than one megabyte in size.
+  JIT model is around one megabyte in size.
 
 - **General**
 
@@ -47,11 +47,11 @@ https://user-images.githubusercontent.com/36505480/144874384-95f80f6d-a4f1-42cc-
 
 - **Flexible sampling rate**
 
-  Silero VAD [supports](https://github.com/snakers4/silero-vad/wiki/Quality-Metrics#sample-rate-comparison)  **8000 Hz** and **16000 Hz** (PyTorch JIT) and **16000 Hz** (ONNX) [sampling rates](https://en.wikipedia.org/wiki/Sampling_(signal_processing)#Sampling_rate).
+  Silero VAD [supports](https://github.com/snakers4/silero-vad/wiki/Quality-Metrics#sample-rate-comparison)  **8000 Hz** and **16000 Hz** [sampling rates](https://en.wikipedia.org/wiki/Sampling_(signal_processing)#Sampling_rate).
 
 - **Flexible chunk size**
 
-  Model was trained on audio chunks of different lengths. **30 ms**, **60 ms** and **100 ms** long chunks are supported directly, others may work as well.
+  Model was trained on **30 ms**. Longer chunks are supported directly, others may work as well.
 
 - **Highly Portable**
 
@@ -105,3 +105,9 @@ Please see our [wiki](https://github.com/snakers4/silero-models/wiki) and [tiers
   email = {hello@silero.ai}
 }
 ```
+
+<br/>
+<h2 align="center">VAD-based Community Apps</h2>
+<br/>
+
+- Voice activity detection for the [browser](https://github.com/ricky0123/vad) using ONNX Runtime Web 

hubconf.py

@@ -15,14 +15,25 @@ from utils_vad import (init_jit_model,
                        OnnxWrapper)
 
 
-def silero_vad(onnx=False):
+def versiontuple(v):
+    return tuple(map(int, (v.split('+')[0].split("."))))
+
+
+def silero_vad(onnx=False, force_onnx_cpu=False):
     """Silero Voice Activity Detector
     Returns a model with a set of utils
     Please see https://github.com/snakers4/silero-vad for usage examples
     """
+
+    if not onnx:
+        installed_version = torch.__version__
+        supported_version = '1.12.0'
+        if versiontuple(installed_version) < versiontuple(supported_version):
+            raise Exception(f'Please install torch {supported_version} or greater ({installed_version} installed)')
+
     hub_dir = torch.hub.get_dir()
     if onnx:
-        model = OnnxWrapper(f'{hub_dir}/snakers4_silero-vad_master/files/silero_vad.onnx')
+        model = OnnxWrapper(f'{hub_dir}/snakers4_silero-vad_master/files/silero_vad.onnx', force_onnx_cpu)
     else:
         model = init_jit_model(model_path=f'{hub_dir}/snakers4_silero-vad_master/files/silero_vad.jit')
     utils = (get_speech_timestamps,
@@ -34,7 +45,7 @@ def silero_vad(onnx=False):
     return model, utils
 
 
-def silero_number_detector(onnx=False):
+def silero_number_detector(onnx=False, force_onnx_cpu=False):
     """Silero Number Detector
     Returns a model with a set of utils
     Please see https://github.com/snakers4/silero-vad for usage examples
@@ -43,7 +54,7 @@ def silero_number_detector(onnx=False):
         url = 'https://models.silero.ai/vad_models/number_detector.onnx'
     else:
         url = 'https://models.silero.ai/vad_models/number_detector.jit'
-    model = Validator(url)
+    model = Validator(url, force_onnx_cpu)
     utils = (get_number_ts,
              save_audio,
              read_audio,
@@ -53,7 +64,7 @@ def silero_number_detector(onnx=False):
     return model, utils
 
 
-def silero_lang_detector(onnx=False):
+def silero_lang_detector(onnx=False, force_onnx_cpu=False):
     """Silero Language Classifier
     Returns a model with a set of utils
     Please see https://github.com/snakers4/silero-vad for usage examples
@@ -62,14 +73,14 @@ def silero_lang_detector(onnx=False):
         url = 'https://models.silero.ai/vad_models/number_detector.onnx'
     else:
         url = 'https://models.silero.ai/vad_models/number_detector.jit'
-    model = Validator(url)
+    model = Validator(url, force_onnx_cpu)
     utils = (get_language,
              read_audio)
 
     return model, utils
 
 
-def silero_lang_detector_95(onnx=False):
+def silero_lang_detector_95(onnx=False, force_onnx_cpu=False):
     """Silero Language Classifier (95 languages)
     Returns a model with a set of utils
     Please see https://github.com/snakers4/silero-vad for usage examples
@@ -80,7 +91,7 @@ def silero_lang_detector_95(onnx=False):
         url = 'https://models.silero.ai/vad_models/lang_classifier_95.onnx'
     else:
         url = 'https://models.silero.ai/vad_models/lang_classifier_95.jit'
-    model = Validator(url)
+    model = Validator(url, force_onnx_cpu)
 
     with open(f'{hub_dir}/snakers4_silero-vad_master/files/lang_dict_95.json', 'r') as f:
         lang_dict = json.load(f)

utils_vad.py

@@ -9,21 +9,21 @@ languages = ['ru', 'en', 'de', 'es']
 
 class OnnxWrapper():
 
-    def __init__(self, path):
+    def __init__(self, path, force_onnx_cpu=False):
         import numpy as np
         global np
         import onnxruntime
-        self.session = onnxruntime.InferenceSession(path)
+        if force_onnx_cpu and 'CPUExecutionProvider' in onnxruntime.get_available_providers():
+            self.session = onnxruntime.InferenceSession(path, providers=['CPUExecutionProvider'])
+        else:
+            self.session = onnxruntime.InferenceSession(path)
         self.session.intra_op_num_threads = 1
         self.session.inter_op_num_threads = 1
 
         self.reset_states()
+        self.sample_rates = [8000, 16000]
 
-    def reset_states(self):
+    def _validate_input(self, x, sr: int):
-        self._h = np.zeros((2, 1, 64)).astype('float32')
-        self._c = np.zeros((2, 1, 64)).astype('float32')
-
-    def __call__(self, x, sr: int):
         if x.dim() == 1:
             x = x.unsqueeze(0)
         if x.dim() > 2:
@@ -34,31 +34,73 @@ class OnnxWrapper():
             x = x[::step]
             sr = 16000
 
-        if x.shape[0] > 1:
+        if sr not in self.sample_rates:
-            raise ValueError("Onnx model does not support batching")
+            raise ValueError(f"Supported sampling rates: {self.sample_rates} (or multiply of 16000)")
-
-        if sr not in [16000]:
-            raise ValueError(f"Supported sample rates: {[16000]}")
 
         if sr / x.shape[1] > 31.25:
             raise ValueError("Input audio chunk is too short")
 
-        ort_inputs = {'input': x.numpy(), 'h0': self._h, 'c0': self._c}
+        return x, sr
-        ort_outs = self.session.run(None, ort_inputs)
-        out, self._h, self._c = ort_outs
 
-        out = torch.tensor(out).squeeze(2)[:, 1]  # make output type match JIT analog
+    def reset_states(self, batch_size=1):
+        self._h = np.zeros((2, batch_size, 64)).astype('float32')
+        self._c = np.zeros((2, batch_size, 64)).astype('float32')
+        self._last_sr = 0
+        self._last_batch_size = 0
 
+    def __call__(self, x, sr: int):
+
+        x, sr = self._validate_input(x, sr)
+        batch_size = x.shape[0]
+
+        if not self._last_batch_size:
+            self.reset_states(batch_size)
+        if (self._last_sr) and (self._last_sr != sr):
+            self.reset_states(batch_size)
+        if (self._last_batch_size) and (self._last_batch_size != batch_size):
+            self.reset_states(batch_size)
+
+        if sr in [8000, 16000]:
+            ort_inputs = {'input': x.numpy(), 'h': self._h, 'c': self._c, 'sr': np.array(sr)}
+            ort_outs = self.session.run(None, ort_inputs)
+            out, self._h, self._c = ort_outs
+        else:
+            raise ValueError()
+
+        self._last_sr = sr
+        self._last_batch_size = batch_size
+
+        out = torch.tensor(out)
         return out
 
+    def audio_forward(self, x, sr: int, num_samples: int = 512):
+        outs = []
+        x, sr = self._validate_input(x, sr)
+
+        if x.shape[1] % num_samples:
+            pad_num = num_samples - (x.shape[1] % num_samples)
+            x = torch.nn.functional.pad(x, (0, pad_num), 'constant', value=0.0)
+
+        self.reset_states(x.shape[0])
+        for i in range(0, x.shape[1], num_samples):
+            wavs_batch = x[:, i:i+num_samples]
+            out_chunk = self.__call__(wavs_batch, sr)
+            outs.append(out_chunk)
+
+        stacked = torch.cat(outs, dim=1)
+        return stacked.cpu()
+
 
 class Validator():
-    def __init__(self, url):
+    def __init__(self, url, force_onnx_cpu):
         self.onnx = True if url.endswith('.onnx') else False
         torch.hub.download_url_to_file(url, 'inf.model')
         if self.onnx:
             import onnxruntime
-            self.model = onnxruntime.InferenceSession('inf.model')
+            if force_onnx_cpu and 'CPUExecutionProvider' in onnxruntime.get_available_providers():
+                self.model = onnxruntime.InferenceSession('inf.model', providers=['CPUExecutionProvider'])
+            else:
+                self.model = onnxruntime.InferenceSession('inf.model')
         else:
             self.model = init_jit_model(model_path='inf.model')
 
@@ -123,7 +165,7 @@ def get_speech_timestamps(audio: torch.Tensor,
                           min_speech_duration_ms: int = 250,
                           max_speech_duration_s: float = float('inf'),
                           min_silence_duration_ms: int = 100,
-                          window_size_samples: int = 1536,
+                          window_size_samples: int = 512,
                           speech_pad_ms: int = 30,
                           return_seconds: bool = False,
                           visualize_probs: bool = False):