app.py

import os
import time
import gradio as gr
import soundfile
import torch

import infer_tool

convert_cnt = [0]
dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# model_name = "83_epochs.pth"
model_name = "mg_1255_epochs_v0.0.6.pth"
config_name = "milky_green.json"
net_g_ms, hubert_soft, feature_input, hps_ms = infer_tool.load_model(f"{model_name}", f"configs/{config_name}")

# 获取config参数
target_sample = hps_ms.data.sampling_rate
spk_dict = {
    "明前奶绿": 0,
    "云灏": 2,
    "即霜": 3,
    "奕兰秋": 4
}

def list_models():
  global model_name
  res = []
  dir = os.getcwd()
  for f in os.listdir(dir):
    if(f.startswith("D_")):
      continue
    if(f.endswith(".pth")):
      res.append(f)
      if len(f) >= len(model_name):
        model_name = f
  return res

def vc_fn(sid, audio_record, audio_upload, tran):
    print(sid, time.strftime("%Y-%m-%d %H:%M:%S", time.localtime()))
    if audio_upload is not None:
        audio_path = audio_upload
    elif audio_record is not None:
        audio_path = audio_record
    else:
        return "你需要上传wav文件或使用网页内置的录音！", None

    audio, sampling_rate = infer_tool.format_wav(audio_path, target_sample)
    duration = audio.shape[0] / sampling_rate
    if duration > 600:
        return "请上传小于600s的音频，需要转换长音频请使用colab", None

    o_audio, out_sr = infer_tool.infer(audio_path, spk_dict[sid], tran, net_g_ms, hubert_soft, feature_input)
    out_path = f"./out_temp.wav"
    soundfile.write(out_path, o_audio, target_sample)
    infer_tool.f0_plt(audio_path, out_path, tran, hubert_soft, feature_input)
    mistake, var = infer_tool.calc_error(audio_path, out_path, tran, feature_input)
    return f"半音偏差：{mistake}\n半音方差：{var}", (
        target_sample, o_audio), gr.Image.update("temp.jpg")

def change_model(model):
  global model_name
  global net_g_ms
  global hubert_soft
  global feature_input
  global hps_ms
  model_name = model
  net_g_ms, hubert_soft, feature_input, hps_ms = infer_tool.load_model(f"{model_name}", f"configs/{config_name}")
  return "载入模型："+model_name

available_models = list_models()

app = gr.Blocks()
with app:
    with gr.Tabs():
        with gr.TabItem("Basic"):
            gr.Markdown(value="""
            本模型为 [sovits_f0](https://huggingface.co/spaces/innnky/nyaru-svc2.0-advanced) 魔改。含AI奶绿（aka. [明前奶绿](https://space.bilibili.com/2132180406)）音色，支持**60s以内**的**无伴奏**wav、mp3（单声道）格式，或使用**网页内置**的录音（二选一）

            转换效果取决于源音频语气、节奏是否与目标音色相近，以及音域是否超出目标音色音域范围

            奶绿高音数据效果稍差，一些音高过高的需要考虑降调

            该模型基于 [innnky/so-vits-svc](https://github.com/innnky/so-vits-svc)，如果想自己制作并训练模型可以访问这个 [GitHub 仓库](https://github.com/IceKyrin/sovits_guide)

            """)
            model_selected = gr.Dropdown(choices=available_models, label = "模型", value=model_name)
            current_model_text = gr.Markdown("")
            model_selected.change(change_model, inputs=[model_selected], outputs=[current_model_text])
            speaker_id = gr.Dropdown(label="音色", choices=['明前奶绿'], value="明前奶绿")
            record_input = gr.Audio(source="microphone", label="录制你的声音", type="filepath", elem_id="audio_inputs")
            upload_input = gr.Audio(source="upload", label="上传音频（长度小于60秒）", type="filepath",
                                    elem_id="audio_inputs")
            vc_transform = gr.Number(label="升降半音（整数，可以正负，半音数量，升高八度就是12）", value=0)
            vc_submit = gr.Button("转换", variant="primary")
            out_audio = gr.Audio(label="Output Audio")
            gr.Markdown(value="""
                        输出信息为音高平均偏差半音数量，体现转换音频的跑调情况（一般平均小于0.5个半音）
                        """)
            out_message = gr.Textbox(label="跑调误差信息")
            gr.Markdown(value="""f0曲线可以直观的显示跑调情况，蓝色为输入音高，橙色为合成音频的音高

                        若**只看见橙色**，说明蓝色曲线被覆盖，转换效果较好

                        """)
            f0_image = gr.Image(label="f0曲线")
        vc_submit.click(vc_fn, [speaker_id, record_input, upload_input, vc_transform],
                        [out_message, out_audio, f0_image])
        with gr.TabItem("使用说明"):
            gr.Markdown(value="""
                        0、合集：https://github.com/IceKyrin/sovits_guide/blob/main/README.md

                        1、仅支持sovit_f0（sovits2.0）模型

                        2、自行下载hubert-soft-0d54a1f4.pt改名为hubert.pt放置于pth文件夹下（已经下好了）
                            https://github.com/bshall/hubert/releases/tag/v0.1

                        3、pth文件夹下放置sovits2.0的模型

                        4、与模型配套的xxx.json，需有speaker项——人物列表

                        5、放无伴奏的音频、或网页内置录音，不要放奇奇怪怪的格式

                        6、仅供交流使用，不对用户行为负责

                        """)

    app.launch()