rvc.py

import os
import glob
import json
import traceback
import logging
import gradio as gr
import numpy as np
import librosa
import torch
import asyncio
import edge_tts
import yt_dlp
import ffmpeg
import subprocess
import sys
import io
import wave
from datetime import datetime
from fairseq import checkpoint_utils
from lib.infer_pack.models import (
    SynthesizerTrnMs256NSFsid,
    SynthesizerTrnMs256NSFsid_nono,
    SynthesizerTrnMs768NSFsid,
    SynthesizerTrnMs768NSFsid_nono,
)
from vc_infer_pipeline import VC
from config import Config
config = Config()
logging.getLogger("numba").setLevel(logging.WARNING)
limitation = os.getenv("SYSTEM") == "spaces"

audio_mode = []
f0method_mode = []
f0method_info = ""
if limitation is True:
    audio_mode = ["Upload audio", "TTS Audio"]
    f0method_mode = ["pm", "crepe", "harvest"]
    f0method_info = "PM is fast, rmvpe is middle, Crepe or harvest is good but it was extremely slow (Default: PM)"
else:
    audio_mode = ["Upload audio", "Youtube", "TTS Audio"]
    f0method_mode = ["pm", "crepe", "harvest"]
    f0method_info = "PM is fast, rmvpe is middle. Crepe or harvest is good but it was extremely slow (Default: PM))"

if os.path.isfile("rmvpe.pt"):
    f0method_mode.insert(2, "rmvpe")

def create_vc_fn(model_title, tgt_sr, net_g, vc, if_f0, version, file_index):
    def vc_fn(
        vc_audio_mode,
        vc_input, 
        vc_upload,
        tts_text,
        tts_voice,
        f0_up_key,
        f0_method,
        index_rate,
        filter_radius,
        resample_sr,
        rms_mix_rate,
        protect,
    ):
        try:
            if vc_audio_mode == "Input path" or "Youtube" and vc_input != "":
                audio, sr = librosa.load(vc_input, sr=16000, mono=True)
            elif vc_audio_mode == "Upload audio":
                if vc_upload is None:
                    return "You need to upload an audio", None
                sampling_rate, audio = vc_upload
                duration = audio.shape[0] / sampling_rate
                if duration > 360 and limitation:
                    return "Please upload an audio file that is less than 1 minute.", 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)
            elif vc_audio_mode == "TTS Audio":
                if len(tts_text) > 600 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)
                vc_input = "tts.mp3"
            times = [0, 0, 0]
            f0_up_key = int(f0_up_key)
            audio_opt = vc.pipeline(
                hubert_model,
                net_g,
                0,
                audio,
                vc_input,
                times,
                f0_up_key,
                f0_method,
                file_index,
                # file_big_npy,
                index_rate,
                if_f0,
                filter_radius,
                tgt_sr,
                resample_sr,
                rms_mix_rate,
                version,
                protect,
                f0_file=None,
            )
            info = f"[{datetime.now().strftime('%Y-%m-%d %H:%M')}]: npy: {times[0]}, f0: {times[1]}s, infer: {times[2]}s"
            print(f"{model_title} | {info}")
            return info, (tgt_sr, audio_opt)
        except:
            info = traceback.format_exc()
            print(info)
            return info, (None, None)
    return vc_fn



def load_model():
    categories = []
    with open("weights/folder_info.json", "r", encoding="utf-8") as f:
        folder_info = json.load(f)
    for category_name, category_info in folder_info.items():
        if not category_info['enable']:
            continue
        category_title = category_info['title']
        category_folder = category_info['folder_path']
        models = []
        with open(f"weights/{category_folder}/model_info.json", "r", encoding="utf-8") as f:
            models_info = json.load(f)
        for character_name, info in models_info.items():
            if not info['enable']:
                continue
            model_title = info['title']
            model_name = info['model_path']
            model_author = info.get("author", None)
            model_cover = f"weights/{category_folder}/{character_name}/{info['cover']}"
            model_index = f"weights/{category_folder}/{character_name}/{info['feature_retrieval_library']}"
            cpt = torch.load(f"weights/{category_folder}/{character_name}/{model_name}", 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)
            version = cpt.get("version", "v1")
            if version == "v1":
                if if_f0 == 1:
                    net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
                else:
                    net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
                model_version = "V1"
            elif version == "v2":
                if if_f0 == 1:
                    net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
                else:
                    net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
                model_version = "V2"
            del net_g.enc_q
            print(net_g.load_state_dict(cpt["weight"], strict=False))
            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)
            print(f"Model loaded: {character_name} / {info['feature_retrieval_library']} | ({model_version})")
            models.append((character_name, model_title, model_author, model_cover, model_version, create_vc_fn(model_title, tgt_sr, net_g, vc, if_f0, version, model_index)))
        categories.append([category_title, category_folder, models])
    return categories



def cut_vocal_and_inst(url, audio_provider, split_model):
    if url != "":
        if not os.path.exists("dl_audio"):
            os.mkdir("dl_audio")
        if audio_provider == "Youtube":
            ydl_opts = {
            'format': 'bestaudio/best',
            'postprocessors': [{
                'key': 'FFmpegExtractAudio',
                'preferredcodec': 'wav',
            }],
            "outtmpl": 'dl_audio/youtube_audio',
            }
            with yt_dlp.YoutubeDL(ydl_opts) as ydl:
                ydl.download([url])
            audio_path = "dl_audio/youtube_audio.wav"
        else:
            # Spotify doesnt work.
            # Need to find other solution soon.
            ''' 
            command = f"spotdl download {url} --output dl_audio/.wav"
            result = subprocess.run(command.split(), stdout=subprocess.PIPE)
            print(result.stdout.decode())
            audio_path = "dl_audio/spotify_audio.wav"
            '''
        if split_model == "htdemucs":
            command = f"demucs --two-stems=vocals {audio_path} -o output"
            result = subprocess.run(command.split(), stdout=subprocess.PIPE)
            print(result.stdout.decode())
            return "output/htdemucs/youtube_audio/vocals.wav", "output/htdemucs/youtube_audio/no_vocals.wav", audio_path, "output/htdemucs/youtube_audio/vocals.wav"
        else:
            command = f"demucs --two-stems=vocals -n mdx_extra_q {audio_path} -o output"
            result = subprocess.run(command.split(), stdout=subprocess.PIPE)
            print(result.stdout.decode())
            return "output/mdx_extra_q/youtube_audio/vocals.wav", "output/mdx_extra_q/youtube_audio/no_vocals.wav", audio_path, "output/mdx_extra_q/youtube_audio/vocals.wav"
    else:
        raise gr.Error("URL Required!")
        return None, None, None, None



def combine_vocal_and_inst(audio_data, audio_volume, split_model):
    if not os.path.exists("output/result"):
        os.mkdir("output/result")
    vocal_path = "output/result/output.wav"
    output_path = "output/result/combine.mp3"
    if split_model == "htdemucs":
        inst_path = "output/htdemucs/youtube_audio/no_vocals.wav"
    else:
        inst_path = "output/mdx_extra_q/youtube_audio/no_vocals.wav"
    with wave.open(vocal_path, "w") as wave_file:
        wave_file.setnchannels(1) 
        wave_file.setsampwidth(2)
        wave_file.setframerate(audio_data[0])
        wave_file.writeframes(audio_data[1].tobytes())
    command =  f'ffmpeg -y -i {inst_path} -i {vocal_path} -filter_complex [1:a]volume={audio_volume}dB[v];[0:a][v]amix=inputs=2:duration=longest -b:a 320k -c:a libmp3lame {output_path}'
    result = subprocess.run(command.split(), stdout=subprocess.PIPE)
    print(result.stdout.decode())
    return output_path

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()