app.py

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 = "1121_epochs.pth"
config_name = "config.json"
net_g_ms, hubert_soft, feature_input, hps_ms = infer_tool.load_model(f"{model_name}", f"{config_name}")

# 获取config参数
target_sample = hps_ms.data.sampling_rate
spk_dict = {
    "鸢一折纸": 0,
    "时崎狂三": 1,
    "冰芽川四糸乃": 2,
    "五河琴里": 3,
    "八舞夕弦": 4,
    "八舞耶俱矢": 5,
    "诱宵美九": 6,
    "夜刀神十香": 7
}


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

    audio, sampling_rate = infer_tool.format_wav(audio_path, target_sample)
    duration = audio.shape[0] / sampling_rate

    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"分段误差参考：0.3优秀，0.5左右合理，少量0.8-1可以接受\n若偏差过大，请调整升降半音数；多次调整均过大、说明超出歌手音域\n半音偏差：{mistake}\n半音方差：{var}", (
        target_sample, o_audio), gr.Image.update("temp.jpg")


app = gr.Blocks()
with app:
    with gr.Tabs():
        with gr.TabItem("Basic"):
            gr.Markdown(value="""
            源码参考: [xiaolang/sovits_f0](https://huggingface.co/spaces/xiaolang/sovits_f0/tree/main)
            
            **音频格式为wav**
            
            转换效果取决于源音频语气、节奏是否与目标音色相近。
            
            源音频为女声时，**建议降3-6key**，**最后的输出误差越接近0，音准越高**
            
            源音频为**低音男声**时，**建议升3key，具体看曲线图情况**
            
            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
            
            """)
            speaker_id = gr.Dropdown(label="音色", choices=list(spk_dict.keys()), value=list(spk_dict.values()))
            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180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_message = gr.Textbox(label="Output Message")
            out_audio = gr.Audio(label="Output Audio")
            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])
    app.launch(share=True)