Merge pull request #56 from iflamed/fastapi

Add Fastapi server to serve TTS and download script

requirements.txt

@@ -26,4 +26,6 @@ tensorboard==2.14.0
 torch==2.0.1
 torchaudio==2.0.2
 wget==3.2
+fastapi==0.111.0
+fastapi-cli==0.0.4
 WeTextProcessing==1.0.3

runtime/python/fastapi_client.py

@@ -0,0 +1,78 @@
+import argparse
+import logging
+import requests
+
+def saveResponse(path, response):
+    # 以二进制写入模式打开文件
+    with open(path, 'wb') as file:
+        # 将响应的二进制内容写入文件
+        file.write(response.content)
+
+def main():
+    api = args.api_base
+    if args.mode == 'sft':
+        url = api + "/api/inference/sft"
+        payload={
+            'tts': args.tts_text,
+            'role': args.spk_id
+        }
+        response = requests.request("POST", url, data=payload)
+        saveResponse(args.tts_wav, response)
+    elif args.mode == 'zero_shot':
+        url = api + "/api/inference/zero-shot"
+        payload={
+            'tts': args.tts_text,
+            'prompt': args.prompt_text
+        }
+        files=[('audio', ('prompt_audio.wav', open(args.prompt_wav,'rb'), 'application/octet-stream'))]
+        response = requests.request("POST", url, data=payload, files=files)
+        saveResponse(args.tts_wav, response)
+    elif args.mode == 'cross_lingual':
+        url = api + "/api/inference/cross-lingual"
+        payload={
+            'tts': args.tts_text,
+        }
+        files=[('audio', ('prompt_audio.wav', open(args.prompt_wav,'rb'), 'application/octet-stream'))]
+        response = requests.request("POST", url, data=payload, files=files)
+        saveResponse(args.tts_wav, response)
+    else:
+        url = api + "/api/inference/instruct"
+        payload = {
+            'tts': args.tts_text,
+            'role': args.spk_id,
+            'instruct': args.instruct_text
+        }
+        response = requests.request("POST", url, data=payload)
+        saveResponse(args.tts_wav, response)
+    logging.info("Response save to {}", args.tts_wav)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--api_base',
+                        type=str,
+                        default='http://127.0.0.1:6006')
+    parser.add_argument('--mode',
+                        default='sft',
+                        choices=['sft', 'zero_shot', 'cross_lingual', 'instruct'],
+                        help='request mode')
+    parser.add_argument('--tts_text',
+                        type=str,
+                        default='你好，我是通义千问语音合成大模型，请问有什么可以帮您的吗？')
+    parser.add_argument('--spk_id',
+                        type=str,
+                        default='中文女')
+    parser.add_argument('--prompt_text',
+                        type=str,
+                        default='希望你以后能够做的比我还好呦。')
+    parser.add_argument('--prompt_wav',
+                        type=str,
+                        default='../../zero_shot_prompt.wav')
+    parser.add_argument('--instruct_text',
+                        type=str,
+                        default='Theo \'Crimson\', is a fiery, passionate rebel leader. Fights with fervor for justice, but struggles with impulsiveness.')
+    parser.add_argument('--tts_wav',
+                        type=str,
+                        default='demo.wav')
+    args = parser.parse_args()
+    prompt_sr, target_sr = 16000, 22050
+    main()

runtime/python/fastapi_server.py

@@ -0,0 +1,109 @@
+# Set inference model
+# export MODEL_DIR=pretrained_models/CosyVoice-300M-Instruct
+# For development
+# fastapi dev --port 6006 fastapi_server.py 
+# For production deployment
+# fastapi run --port 6006 fastapi_server.py 
+
+import os
+import sys
+import io,time
+from fastapi import FastAPI, Response, File, UploadFile, Form
+from fastapi.responses import HTMLResponse
+from contextlib import asynccontextmanager
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/../..'.format(ROOT_DIR))
+sys.path.append('{}/../../third_party/Matcha-TTS'.format(ROOT_DIR))
+from cosyvoice.cli.cosyvoice import CosyVoice
+from cosyvoice.utils.file_utils import load_wav
+import numpy as np
+import torch
+import torchaudio
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+
+class LaunchFailed(Exception):
+    pass
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    model_dir = os.getenv("MODEL_DIR", "pretrained_models/CosyVoice-300M-SFT")
+    if model_dir:
+        logging.info("MODEL_DIR is {}", model_dir)
+        app.cosyvoice = CosyVoice('../../'+model_dir)
+        # sft usage
+        logging.info("Avaliable speakers {}", app.cosyvoice.list_avaliable_spks())
+    else:
+        raise LaunchFailed("MODEL_DIR environment must set")
+    yield
+
+app = FastAPI(lifespan=lifespan)
+
+def buildResponse(output):
+    buffer = io.BytesIO()
+    torchaudio.save(buffer, output, 22050, format="wav")
+    buffer.seek(0)
+    return Response(content=buffer.read(-1), media_type="audio/wav")
+
+@app.post("/api/inference/sft")
+@app.get("/api/inference/sft")
+async def sft(tts: str = Form(), role: str = Form()):
+    start = time.process_time()
+    output = app.cosyvoice.inference_sft(tts, role)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.post("/api/inference/zero-shot")
+async def zeroShot(tts: str = Form(), prompt: str = Form(), audio: UploadFile = File()):
+    start = time.process_time()
+    prompt_speech = load_wav(audio.file, 16000)
+    prompt_audio = (prompt_speech.numpy() * (2**15)).astype(np.int16).tobytes()
+    prompt_speech_16k = torch.from_numpy(np.array(np.frombuffer(prompt_audio, dtype=np.int16))).unsqueeze(dim=0)
+    prompt_speech_16k = prompt_speech_16k.float() / (2**15)
+
+    output = app.cosyvoice.inference_zero_shot(tts, prompt, prompt_speech_16k)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.post("/api/inference/cross-lingual")
+async def crossLingual(tts: str = Form(), audio: UploadFile = File()):
+    start = time.process_time()
+    prompt_speech = load_wav(audio.file, 16000)
+    prompt_audio = (prompt_speech.numpy() * (2**15)).astype(np.int16).tobytes()
+    prompt_speech_16k = torch.from_numpy(np.array(np.frombuffer(prompt_audio, dtype=np.int16))).unsqueeze(dim=0)
+    prompt_speech_16k = prompt_speech_16k.float() / (2**15)
+
+    output = app.cosyvoice.inference_cross_lingual(tts, prompt_speech_16k)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.post("/api/inference/instruct")
+@app.get("/api/inference/instruct")
+async def instruct(tts: str = Form(), role: str = Form(), instruct: str = Form()):
+    start = time.process_time()
+    output = app.cosyvoice.inference_instruct(tts, role, instruct)
+    end = time.process_time()
+    logging.info("infer time is {} seconds", end-start)
+    return buildResponse(output['tts_speech'])
+
+@app.get("/api/roles")
+async def roles():
+    return {"roles": app.cosyvoice.list_avaliable_spks()}
+
+@app.get("/", response_class=HTMLResponse)
+async def root():
+    return """
+    <!DOCTYPE html>
+    <html lang=zh-cn>
+        <head>
+            <meta charset=utf-8>
+            <title>Api information</title>
+        </head>
+        <body>
+            Get the supported tones from the Roles API first, then enter the tones and textual content in the TTS API for synthesis. <a href='./docs'>Documents of API</a>
+        </body>
+    </html>
+    """