diff --git a/main.py b/main.py
deleted file mode 100644
index d212dd3..0000000
--- a/main.py
+++ /dev/null
@@ -1,40 +0,0 @@
-import io,time
-from fastapi import FastAPI, Response
-from fastapi.responses import HTMLResponse
-from cosyvoice.cli.cosyvoice import CosyVoice
-import torchaudio
-
-cosyvoice = CosyVoice('pretrained_models/CosyVoice-300M-SFT')
-# sft usage
-print(cosyvoice.list_avaliable_spks())
-app = FastAPI()
-
-@app.get("/api/voice/tts")
-async def tts(query: str, role: str):
- start = time.process_time()
- output = cosyvoice.inference_sft(query, role)
- end = time.process_time()
- print("infer time:", end-start, "seconds")
- buffer = io.BytesIO()
- torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
- buffer.seek(0)
- return Response(content=buffer.read(-1), media_type="audio/wav")
-
-@app.get("/api/voice/roles")
-async def roles():
- return {"roles": cosyvoice.list_avaliable_spks()}
-
-@app.get("/", response_class=HTMLResponse)
-async def root():
- return """
-
-
-
-
- Api information
-
-
- Get the supported tones from the Roles API first, then enter the tones and textual content in the TTS API for synthesis. Documents of API
-
-
- """
diff --git a/runtime/python/fastapi_server.py b/runtime/python/fastapi_server.py
new file mode 100644
index 0000000..f718373
--- /dev/null
+++ b/runtime/python/fastapi_server.py
@@ -0,0 +1,102 @@
+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 """
+
+
+
+
+ Api information
+
+
+ Get the supported tones from the Roles API first, then enter the tones and textual content in the TTS API for synthesis. Documents of API
+
+
+ """