api.py

import torch
from fastapi import FastAPI
from pydantic import BaseModel
import numpy as np
import base64
import io
from scipy.io.wavfile import write
import sounddevice as sd

# 自定义模块
import commons
import utils
from models import SynthesizerTrn
from text.symbols import symbols
from text import text_to_sequence

# 检查 PyTorch 版本
print(torch.__version__)

# 检查 CUDA 是否可用
print(torch.cuda.is_available())

# 检查当前 CUDA 版本
print(torch.version.cuda)

# FastAPI 应用
app = FastAPI()

# 请求体模型
class TextRequest(BaseModel):
    text: str

# 加载配置和模型
config_path = "configs/steins_gate_base.json"
checkpoint_path = "G_265000.pth"
hps = utils.get_hparams_from_file(config_path)
net_g = SynthesizerTrn(
    len(symbols),
    hps.data.filter_length // 2 + 1,
    hps.train.segment_size // hps.data.hop_length,
    **hps.model,
).eval()
utils.load_checkpoint(checkpoint_path, net_g, None)

# 文本到语音合成
def text_to_speech(content):
    stn_tst = text_to_sequence(content, hps.data.text_cleaners)
    if hps.data.add_blank:
        stn_tst = commons.intersperse(stn_tst, 0)
    stn_tst = torch.LongTensor(stn_tst)
    with torch.no_grad():
        x_tst = stn_tst.unsqueeze(0)
        x_tst_lengths = torch.LongTensor([stn_tst.size(0)])
        audio = net_g.infer(x_tst, x_tst_lengths, noise_scale=0.667, noise_scale_w=0.8, length_scale=1)[0][0, 0].data.float().numpy()

    return hps.data.sampling_rate, audio

# API 路由：文本转语音
@app.post("/synthesize")
def synthesize(request: TextRequest):
    # 假设 text_to_speech 是生成音频的函数
    sampling_rate, audio = text_to_speech(request.text)

    # 将音频数据保存到 BytesIO 对象
    wav_bytes = io.BytesIO()
    write(wav_bytes, sampling_rate, (audio * 32767).astype(np.int16))
    wav_bytes.seek(0)  # 将指针移动到文件开头

    # 将 WAV 文件编码为 Base64
    audio_base64 = base64.b64encode(wav_bytes.read()).decode("utf-8")
    return {"audio": audio_base64}

# 主函数
if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="127.0.0.1", port=8000)