chilge/nemo

1import io2import logging3 4import soundfile5import torch6import torchaudio7from flask import Flask, request, send_file8from flask_cors import CORS9 10from inference.infer_tool import Svc, RealTimeVC11 12app = Flask(__name__)13 14CORS(app)15 16logging.getLogger('numba').setLevel(logging.WARNING)17 18 19@app.route("/voiceChangeModel", methods=["POST"])20def voice_change_model():21    request_form = request.form22    wave_file = request.files.get("sample", None)23    # 变调信息24    f_pitch_change = float(request_form.get("fPitchChange", 0))25    # DAW所需的采样率26    daw_sample = int(float(request_form.get("sampleRate", 0)))27    speaker_id = int(float(request_form.get("sSpeakId", 0)))28    # http获得wav文件并转换29    input_wav_path = io.BytesIO(wave_file.read())30 31    # 模型推理32    if raw_infer:33        out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)34        tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample)35    else:36        out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path)37        tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample)38    # 返回音频39    out_wav_path = io.BytesIO()40    soundfile.write(out_wav_path, tar_audio.cpu().numpy(), daw_sample, format="wav")41    out_wav_path.seek(0)42    return send_file(out_wav_path, download_name="temp.wav", as_attachment=True)43 44 45if __name__ == '__main__':46    # 启用则为直接切片合成，False为交叉淡化方式47    # vst插件调整0.3-0.5s切片时间可以降低延迟，直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音48    # 自行选择能接受的方法，或将vst最大切片时间调整为1s，此处设为Ture，延迟大音质稳定一些49    raw_infer = True50    # 每个模型和config是唯一对应的51    model_name = "logs/32k/G_174000-Copy1.pth"52    config_name = "configs/config.json"53    svc_model = Svc(model_name, config_name)54    svc = RealTimeVC()55    # 此处与vst插件对应，不建议更改56    app.run(port=6842, host="0.0.0.0", debug=False, threaded=False)57