anstonjie commented 2 weeks ago

这个api怎么调用？还是要自己封装

aluminumbox commented 2 weeks ago

we have provided grpc example in runtime, or you can wrap your own api

iflamed commented 2 weeks ago

Try this example:

Install fastapi first: pip install fastapi, then fastapi dev --port 3001

import io,time
from fastapi import FastAPI, Response
from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav
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):
    start = time.process_time()
    output = cosyvoice.inference_sft(query, '中文女')
    end = time.process_time()
    print("infer time:", end-start)
    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("/")
async def root():
    return {"message": "Hello World"}

aluminumbox commented 2 weeks ago

Try this example:

Install fastapi first: pip install fastapi, then fastapi dev --port 3001

import io,time
from fastapi import FastAPI, Response
from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav
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):
    start = time.process_time()
    output = cosyvoice.inference_sft(query, '中文女')
    end = time.process_time()
    print("infer time:", end-start)
    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("/")
async def root():
    return {"message": "Hello World"}

we will be happy if you can make a fastapi pr

iflamed commented 2 weeks ago

Try this example: Install fastapi first: pip install fastapi, then fastapi dev --port 3001

import io,time
from fastapi import FastAPI, Response
from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav
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):
    start = time.process_time()
    output = cosyvoice.inference_sft(query, '中文女')
    end = time.process_time()
    print("infer time:", end-start)
    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("/")
async def root():
    return {"message": "Hello World"}

we will be happy if you can make a fastapi pr

Ok， let me create PR.

anstonjie commented 2 weeks ago

thank you

QIN2DIM commented 2 weeks ago

@iflamed why not gradio_client ...

sdrz001 commented 2 weeks ago

能制作个API吗

lonngxiang commented 2 weeks ago

同求，看官方提供的grpc api使用有些复杂

lonngxiang commented 2 weeks ago

Try this example:

Install fastapi first: pip install fastapi, then fastapi dev --port 3001

import io,time
from fastapi import FastAPI, Response
from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav
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):
    start = time.process_time()
    output = cosyvoice.inference_sft(query, '中文女')
    end = time.process_time()
    print("infer time:", end-start)
    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("/")
async def root():
    return {"message": "Hello World"}

如果是inference_zero_shot需要上传音频的需要怎么封装呢

iflamed commented 2 weeks ago

我也不会呀，你会python不？Sent from newbing’s iPhoneOn Jul 10, 2024, at 11:56, dragon10 @.***> wrote:

Try this example: Install fastapi first: pip install fastapi, then fastapi dev --port 3001 import io,time from fastapi import FastAPI, Response from cosyvoice.cli.cosyvoice import CosyVoice from cosyvoice.utils.file_utils import load_wav 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): start = time.process_time() output = cosyvoice.inference_sft(query, '中文女') end = time.process_time() print("infer time:", end-start) 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("/") async def root(): return {"message": "Hello World"}

如果是inference_zero_shot需要上传音频的需要怎么封装呢

—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: @.***>

lonngxiang commented 2 weeks ago

已封装，供参考（代码说明可以参考 https://blog.csdn.net/weixin_42357472/article/details/140321056）：

api.py


import time
import io, os, sys
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))

import numpy as np
from flask import Flask, request, Response
import torch
import torchaudio

from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav

##指定模型地址,对于零样本/跨语言推理，请使用CosyVoice-300M模型。对于SFT推理，
##请使用CosyVoice-300M-SFT模型。对于指令推理，请使用CosyVoice-300M-Instruct模型。
cosyvoice = CosyVoice('/data/pretrained_models/CosyVoice-300M') 

print(cosyvoice.list_avaliable_spks())

app = Flask(__name__)

@app.route("/inference/sft", methods=['POST'])
def sft():
    question_data = request.get_json()
    query = question_data.get('query')
    speaker = question_data.get('speaker')
    if not query:
        return {"error": "Query parameter 'query' is required"}, 400

    start = time.process_time()
    output = cosyvoice.inference_sft(query, speaker)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

@app.route("/inference/zero_shot", methods=['POST'])
def zero_shot():
    question_data = request.get_json()

    tts_text = question_data.get('query')
    prompt_text = question_data.get('prompt_text')
    prompt_speech = load_wav(question_data.get('prompt_speech'), 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)

    start = time.process_time()
    output = cosyvoice.inference_zero_shot(tts_text, prompt_text, prompt_speech_16k)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

@app.route("/inference/cross_lingual", methods=['POST'])
def cross_lingual():
    question_data = request.get_json()

    tts_text = question_data.get('query')
    prompt_speech = load_wav(question_data.get('prompt_speech'), 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)

    start = time.process_time()
    output = cosyvoice.inference_cross_lingual(tts_text, prompt_speech_16k)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

@app.route("/inference/instruct", methods=['POST'])
def instruct():
    question_data = request.get_json()
    tts_text = question_data.get('query')
    speaker = question_data.get('speaker')
    instruct_text = question_data.get('instruct_text')

    start = time.process_time()
    output = cosyvoice.inference_instruct(tts_text, speaker, instruct_text)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

if __name__ == "__main__":
    app.run(host='0.0.0.0', port=50000,)

lonngxiang commented 2 weeks ago

已封装，供参考，代码说明可以参考 https://blog.csdn.net/weixin_42357472/article/details/140321056

api.py


import time
import io, os, sys
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))

import numpy as np
from flask import Flask, request, Response
import torch
import torchaudio

from cosyvoice.cli.cosyvoice import CosyVoice
from cosyvoice.utils.file_utils import load_wav

##指定模型地址,对于零样本/跨语言推理，请使用CosyVoice-300M模型。对于SFT推理，
##请使用CosyVoice-300M-SFT模型。对于指令推理，请使用CosyVoice-300M-Instruct模型。
cosyvoice = CosyVoice('/data/pretrained_models/CosyVoice-300M') 

print(cosyvoice.list_avaliable_spks())

app = Flask(__name__)

@app.route("/inference/sft", methods=['POST'])
def sft():
    question_data = request.get_json()
    query = question_data.get('query')
    speaker = question_data.get('speaker')
    if not query:
        return {"error": "Query parameter 'query' is required"}, 400

    start = time.process_time()
    output = cosyvoice.inference_sft(query, speaker)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

@app.route("/inference/zero_shot", methods=['POST'])
def zero_shot():
    question_data = request.get_json()

    tts_text = question_data.get('query')
    prompt_text = question_data.get('prompt_text')
    prompt_speech = load_wav(question_data.get('prompt_speech'), 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)

    start = time.process_time()
    output = cosyvoice.inference_zero_shot(tts_text, prompt_text, prompt_speech_16k)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

@app.route("/inference/cross_lingual", methods=['POST'])
def cross_lingual():
    question_data = request.get_json()

    tts_text = question_data.get('query')
    prompt_speech = load_wav(question_data.get('prompt_speech'), 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)

    start = time.process_time()
    output = cosyvoice.inference_cross_lingual(tts_text, prompt_speech_16k)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

@app.route("/inference/instruct", methods=['POST'])
def instruct():
    question_data = request.get_json()
    tts_text = question_data.get('query')
    speaker = question_data.get('speaker')
    instruct_text = question_data.get('instruct_text')

    start = time.process_time()
    output = cosyvoice.inference_instruct(tts_text, speaker, instruct_text)
    end = time.process_time()
    print("infer time:", end - start)
    buffer = io.BytesIO()
    torchaudio.save(buffer, output['tts_speech'], 22050, format="wav")
    buffer.seek(0)
    return Response(buffer.read(), mimetype="audio/wav")

if __name__ == "__main__":
    app.run(host='0.0.0.0', port=50000,)

FunAudioLLM / CosyVoice

这个api怎么调用？还是要自己封装 #47

sft usage