app.py

import os
import glob
import json
import argparse
import traceback
import logging
import gradio as gr
import numpy as np
import librosa
import torch
import asyncio
import edge_tts
from datetime import datetime
from fairseq import checkpoint_utils
from infer_pack.models import SynthesizerTrnMs256NSFsid, SynthesizerTrnMs256NSFsid_nono
from vc_infer_pipeline import VC
from config import Config
config = Config()
logging.getLogger("numba").setLevel(logging.WARNING)
limitation = os.getenv("SYSTEM") == "spaces"  # limit audio length in huggingface spaces

def create_vc_fn(tgt_sr, net_g, vc, if_f0, file_index):
    def vc_fn(
        input_audio,
        f0_up_key,
        f0_method,
        index_rate,
        tts_mode,
        tts_text,
        tts_voice
    ):
        try:
            if tts_mode:
                if len(tts_text) > 100 and limitation:
                    return "Text is too long", None
                if tts_text is None or tts_voice is None:
                    return "You need to enter text and select a voice", None
                asyncio.run(edge_tts.Communicate(tts_text, "-".join(tts_voice.split('-')[:-1])).save("tts.mp3"))
                audio, sr = librosa.load("tts.mp3", sr=16000, mono=True)
            else:
                if config.files:
                    audio, sr = librosa.load(input_audio, sr=16000, mono=True)
                else:
                    if input_audio is None:
                        return "You need to upload an audio", None
                    sampling_rate, audio = input_audio
                    duration = audio.shape[0] / sampling_rate
                    if duration > 600 and limitation:
                        return "Please upload an audio file that is less than 600 seconds. If you need to generate a longer audio file, please use Colab.", None
                    audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
                    if len(audio.shape) > 1:
                        audio = librosa.to_mono(audio.transpose(1, 0))
                    if sampling_rate != 16000:
                        audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
            times = [0, 0, 0]
            f0_up_key = int(f0_up_key)
            audio_opt = vc.pipeline(
                hubert_model,
                net_g,
                0,
                audio,
                times,
                f0_up_key,
                f0_method,
                file_index,
                index_rate,
                if_f0,
                f0_file=None,
            )
            print(
                f"[{datetime.now().strftime('%Y-%m-%d %H:%M')}]: npy: {times[0]}, f0: {times[1]}s, infer: {times[2]}s"
            )
            return "Success", (tgt_sr, audio_opt)
        except:
            info = traceback.format_exc()
            print(info)
            return info, (None, None)
    return vc_fn

def load_hubert():
    global hubert_model
    models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
        ["hubert_base.pt"],
        suffix="",
    )
    hubert_model = models[0]
    hubert_model = hubert_model.to(config.device)
    if config.is_half:
        hubert_model = hubert_model.half()
    else:
        hubert_model = hubert_model.float()
    hubert_model.eval()

def change_to_tts_mode(tts_mode):
    if tts_mode:
        return gr.Audio.update(visible=False), gr.Textbox.update(visible=True), gr.Dropdown.update(visible=True)
    else:
        return gr.Audio.update(visible=True), gr.Textbox.update(visible=False), gr.Dropdown.update(visible=False)

if __name__ == '__main__':
    load_hubert()
    models = []
    tts_voice_list = asyncio.get_event_loop().run_until_complete(edge_tts.list_voices())
    voices = [f"{v['ShortName']}-{v['Gender']}" for v in tts_voice_list]
    folder_path = "weights"
    for name in os.listdir(folder_path):
        print("check folder: " + name)
        if name.startswith("."): break
        cover_path = glob.glob(f"{folder_path}/{name}/*.png") + glob.glob(f"{folder_path}/{name}/*.jpg")
        index_path = glob.glob(f"{folder_path}/{name}/*.index")
        checkpoint_path = glob.glob(f"{folder_path}/{name}/*.pth")
        title = name
        cover = cover_path[0]
        index = index_path[0]
        cpt = torch.load(checkpoint_path[0], map_location="cpu")
        tgt_sr = cpt["config"][-1]
        cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]  # n_spk
        if_f0 = cpt.get("f0", 1)
        if if_f0 == 1:
            net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
        else:
            net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
        del net_g.enc_q
        print(net_g.load_state_dict(cpt["weight"], strict=False))  # 不加这一行清不干净, 真奇葩
        net_g.eval().to(config.device)
        if config.is_half:
            net_g = net_g.half()
        else:
            net_g = net_g.float()
        vc = VC(tgt_sr, config)
        models.append((name, title, cover, create_vc_fn(tgt_sr, net_g, vc, if_f0, index)))
    with gr.Blocks() as app:
        gr.Markdown(
            "# <center>🥳🎶🎡 - AI歌手，RVC歌声转换"
            "## <center>Powered by [RVC-Project](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)"
            "### <center>更多精彩应用，敬请关注[滔滔AI](http://www.talktalkai.com)；滔滔AI，为爱滔滔！💕</center>"
        )
        with gr.Tabs():
            for (name, title, cover, vc_fn) in models:
                with gr.TabItem(name):
                    with gr.Row():
                        gr.Markdown(
                            '<div align="center">'
                            f'<div>{title}</div>\n'+
                            (f'<img style="width:auto;height:300px;" src="file/{cover}">' if cover else "")+
                            '</div>'
                        )
                    with gr.Row():
                        with gr.Column():
                            vc_input = gr.Audio(label="Input audio"+' (less than 10 minutes)' if limitation else '')
                            vc_transpose = gr.Number(label="Transpose", value=0)
                            vc_f0method = gr.Radio(
                                label="Pitch extraction algorithm, PM is fast but Harvest is better for low frequencies",
                                choices=["pm", "harvest"],
                                value="pm",
                                interactive=True,
                            )
                            vc_index_ratio = gr.Slider(
                                minimum=0,
                                maximum=1,
                                label="Retrieval feature ratio",
                                value=0.6,
                                interactive=True,
                            )
                            tts_mode = gr.Checkbox(label="tts (use edge-tts as input)", value=False)
                            tts_text = gr.Textbox(visible=False,label="TTS text (100 words limitation)" if limitation else "TTS text")
                            tts_voice = gr.Dropdown(label="Edge-tts speaker", choices=voices, visible=False, allow_custom_value=False, value="en-US-AnaNeural-Female")
                            vc_submit = gr.Button("Generate", variant="primary")
                        with gr.Column():
                            vc_output1 = gr.Textbox(label="Output Message")
                            vc_output2 = gr.Audio(label="Output Audio")
                vc_submit.click(vc_fn, [vc_input, vc_transpose, vc_f0method, vc_index_ratio, tts_mode, tts_text, tts_voice], [vc_output1, vc_output2])
                tts_mode.change(change_to_tts_mode, [tts_mode], [vc_input, tts_text, tts_voice])
        gr.Markdown("### <center>注意❗：请不要生成会对个人以及组织造成侵害的内容，此程序仅供科研、学习及个人娱乐使用。用户生成内容与程序开发者无关，请自觉合法合规使用，违反者一切后果自负。</center>")
        gr.HTML('''
            <div class="footer">
                        <p>🌊🏞️🎶 - 江水东流急，滔滔无尽声。 明·顾璘
                        </p>
            </div>
        ''') 
        app.queue(concurrency_count=1, max_size=20, api_open=config.api).launch(share=config.share, show_error=True)