app.py

print("Starting up. Please be patient...")

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
from edgetts_db import tts_order_voice

#fuck intel
os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"

config = Config()
logging.getLogger("numba").setLevel(logging.WARNING)
limitation = os.getenv("SYSTEM") == "spaces"
#limitation=True
language_dict = tts_order_voice

authors = ["dacoolkid44", "Hijack", "Maki Ligon", "megaaziib", "Kit Lemonfoot", "yeey5", "Sui", "MahdeenSky"]

f0method_mode = []
if limitation is True:
    f0method_info = "PM is better for testing, RMVPE is better for finalized generations. (Default: PM)"
    f0method_mode = ["pm", "rmvpe"]
else:
    f0method_info = "PM is fast but low quality, crepe and harvest are slow but good quality, RMVPE is the best of both worlds. (Default: PM)"
    f0method_mode = ["pm", "crepe", "harvest", "rmvpe"]

#Eagerload VCs
print("Preloading VCs...")
vcArr=[]
vcArr.append(VC(32000, config))
vcArr.append(VC(40000, config))
vcArr.append(VC(48000, config))

def infer(name, path, index, vc_input, vc_upload, tts_text, tts_voice, f0_up_key, f0_method, index_rate, filter_radius, resample_sr, rms_mix_rate, protect):
    try:
        #Setup audio
        audio=None
        #Determine audio mode
        #TTS takes priority over uploads.
        #Uploads takes priority over paths.
        vc_audio_mode = ""
        #Edge-TTS
        if(tts_text):
            vc_audio_mode = "ETTS"
            if len(tts_text) > 250 and limitation:
                return "Text is too long.", None
            if tts_text is None or tts_voice is None or tts_text=="":
                return "You need to enter text and select a voice.", None
            voice = language_dict[tts_voice]
            try:
                asyncio.run(edge_tts.Communicate(tts_text, voice).save("tts.mp3"))
            except:
                print("Failed to get E-TTS handle. A restart may be needed soon.")
                return "ERROR: Failed to communicate with Edge-TTS. The Edge-TTS service may be down or cannot communicate. Please try another method or try again later.", None
            try:
                audio, sr = librosa.load("tts.mp3", sr=16000, mono=True)
            except:
                return "ERROR: Invalid characters for the chosen TTS speaker. (Change your TTS speaker to one that supports your language!)", None
            duration = audio.shape[0] / sr
            if duration > 30 and limitation:
                return "Your text generated an audio that was too long.", None
            vc_input = "tts.mp3"
        #File upload
        elif(vc_upload):
            vc_audio_mode = "Upload"
            sampling_rate, audio = vc_upload
            duration = audio.shape[0] / sampling_rate
            if duration > 60 and limitation:
                return "Too long! 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)
            tts_text = "Uploaded Audio"
        #YouTube or path
        elif(vc_input):
            audio, sr = librosa.load(vc_input, sr=16000, mono=True)
            vc_audio_mode = "YouTube"
            tts_text = "YouTube Audio"
        else:
            return "Please upload or choose some type of audio.", None  

        if audio is None:
            if vc_audio_mode == "ETTS":
                print("Failed to get E-TTS handle. A restart may be needed soon.")
                return "ERROR: Failed to obtain a correct response from Edge-TTS. The Edge-TTS service may be down or unable to communicate. Please try another method or try again later.", None
            return "ERROR: Unknown audio error. Please try again.", None
            
        times = [0, 0, 0]
        f0_up_key = int(f0_up_key)

        #Setup model
        cpt = torch.load(f"{path}", 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()
        vcIdx = int((tgt_sr/8000)-4)

        #Gen audio
        audio_opt = vcArr[vcIdx].pipeline(
            hubert_model,
            net_g,
            0,
            audio,
            vc_input,
            times,
            f0_up_key,
            f0_method,
            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"Successful inference with model {name} | {tts_text} | {info}")
        del net_g, cpt
        return info, (tgt_sr, audio_opt)
    except:
        info = traceback.format_exc()
        print(info)
        return info, (None, None)

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 = []
        print(f"Creating category {category_title}...")
        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']}"
            if info['feature_retrieval_library'] == "None":
                model_index = None
            if model_index:
               assert os.path.exists(model_index), f"Model {model_title} failed to load index."
            if not (model_author in authors or "/" in model_author or "&" in model_author):
                authors.append(model_author)
            model_path =  f"weights/{category_folder}/{character_name}/{model_name}"
            cpt = torch.load(f"weights/{category_folder}/{character_name}/{model_name}", map_location="cpu")
            model_version = cpt.get("version", "v1")
            print(f"Indexed model {model_title} by {model_author} ({model_version})")
            models.append((character_name, model_title, model_author, model_cover, model_version, model_path, model_index))
            del cpt
        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 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()

if __name__ == '__main__':
    load_hubert()
    categories = load_model()
    voices = list(language_dict.keys())

    # Gradio preloading
    # Input and Upload
    vc_upload = gr.Audio(label="Upload or record an audio file", interactive=True)
    # Youtube
    vc_input = gr.Textbox(label="Input audio path", visible=False)
    vc_download_audio = gr.Dropdown(label="Provider", choices=["Youtube"], allow_custom_value=False, value="Youtube", info="Select provider (Default: Youtube)")
    vc_link = gr.Textbox(label="Youtube URL", info="Example: https://www.youtube.com/watch?v=Nc0sB1Bmf-A", placeholder="https://www.youtube.com/watch?v=...")
    vc_split_model = gr.Dropdown(label="Splitter Model", choices=["htdemucs", "mdx_extra_q"], allow_custom_value=False, value="htdemucs", info="Select the splitter model (Default: htdemucs)")
    vc_split = gr.Button("Split Audio", variant="primary")
    vc_vocal_preview = gr.Audio(label="Vocal Preview")
    vc_inst_preview = gr.Audio(label="Instrumental Preview")
    vc_audio_preview = gr.Audio(label="Audio Preview")
    # TTS
    tts_text = gr.Textbox(label="TTS text", info="Text to speech input (There is a limit of 250 characters)", interactive=True)
    tts_voice = gr.Dropdown(label="Edge-TTS speaker", choices=voices, allow_custom_value=False, value="English-Ana (Female)", interactive=True)
    # Other settings
    vc_transform0 = gr.Number(label="Transpose", value=0, info='Type "12" to change from male to female voice. Type "-12" to change female to male voice')
    f0method0 = gr.Radio(
        label="Pitch extraction algorithm",
        info=f0method_info,
        choices=f0method_mode,
        value="pm",
        interactive=True
    )
    index_rate1 = gr.Slider(
        minimum=0,
        maximum=1,
        label="Retrieval feature ratio",
        info="Accent control. Too high will usually sound too robotic. (Default: 0.4)",
        value=0.4,
        interactive=True,
    )
    filter_radius0 = gr.Slider(
        minimum=0,
        maximum=7,
        label="Apply Median Filtering",
        info="The value represents the filter radius and can reduce breathiness.",
        value=1,
        step=1,
        interactive=True,
    )
    resample_sr0 = gr.Slider(
        minimum=0,
        maximum=48000,
        label="Resample the output audio",
        info="Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling.",
        value=0,
        step=1,
        interactive=True,
    )
    rms_mix_rate0 = gr.Slider(
        minimum=0,
        maximum=1,
        label="Volume Envelope",
        info="Use the volume envelope of the input to replace or mix with the volume envelope of the output. The closer the ratio is to 1, the more the output envelope is used",
        value=1,
        interactive=True,
    )
    protect0 = gr.Slider(
        minimum=0,
        maximum=0.5,
        label="Voice Protection",
        info="Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy",
        value=0.23,
        step=0.01,
        interactive=True,
    )

    with gr.Blocks(theme=gr.themes.Base()) as app:
        gr.Markdown(
            "# <center> VTuber RVC Models\n"
            "### <center> Space by Kit Lemonfoot / Noel Shirogane's High Flying Birds"
            "<center> Original space by megaaziib & zomehwh\n"
            "### <center> Please credit the original model authors if you use this Space."
            "<center>Do no evil.\n\n"
            "[![image](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1Til3SY7-X0x3Wss3YXlgfq8go39DzWHk)\n\n"
        )
        gr.Markdown("<center> Looking for more models? <a href=\"https://docs.google.com/spreadsheets/d/1tvZSggOsZGAPjbMrWOAAaoJJFpJuQlwUEQCf5x1ssO8\">Check out the VTuber AI Model Tracking spreadsheet!</a>")
        for (folder_title, folder, models) in categories:
            with gr.TabItem(folder_title):
                with gr.Tabs():
                    if not models:
                        gr.Markdown("# <center> No Model Loaded.")
                        gr.Markdown("## <center> Please add model or fix your model path.")
                        continue
                    for (name, title, author, cover, model_version, model_path, model_index) in models:
                        with gr.TabItem(name):
                            with gr.Row():
                                with gr.Column():
                                    gr.Markdown(
                                        '<div align="center">'
                                        f'<div>{title}</div>\n'+
                                        f'<div>RVC {model_version} Model</div>\n'+
                                        (f'<div>Model author: {author}</div>' if author else "")+
                                        (f'<img style="width:auto;height:300px;" src="file/{cover}"></img>' if cover else "")+
                                        '</div>'
                                    )
                                with gr.Column():
                                    vc_log = gr.Textbox(label="Output Information", interactive=False)
                                    vc_output = gr.Audio(label="Output Audio", interactive=False)
                                    #This is a fucking stupid solution but Gradio refuses to pass in values unless I do this.
                                    vc_name = gr.Textbox(value=title, visible=False, interactive=False)
                                    vc_mp = gr.Textbox(value=model_path, visible=False, interactive=False)
                                    vc_mi = gr.Textbox(value=model_index, visible=False, interactive=False)
                                    vc_convert = gr.Button("Convert", variant="primary")

                                    vc_convert.click(
                                        fn=infer,
                                        inputs=[
                                            vc_name,
                                            vc_mp,
                                            vc_mi,
                                            vc_input, 
                                            vc_upload,
                                            tts_text,
                                            tts_voice,
                                            vc_transform0,
                                            f0method0,
                                            index_rate1,
                                            filter_radius0,
                                            resample_sr0,
                                            rms_mix_rate0,
                                            protect0
                                        ],
                                        outputs=[vc_log, vc_output]
                                    )

        with gr.Row():
            with gr.Column():
                with gr.Tab("Edge-TTS"):
                    tts_text.render()
                    tts_voice.render()
                with gr.Tab("Upload/Record"):
                    vc_input.render()
                    vc_upload.render()
                if(not limitation):
                    with gr.Tab("YouTube"):
                        vc_download_audio.render()
                        vc_link.render()
                        vc_split_model.render()
                        vc_split.render()
                        vc_vocal_preview.render()
                        vc_inst_preview.render()
                        vc_audio_preview.render()
            with gr.Column():
                vc_transform0.render()
                f0method0.render()
                index_rate1.render()
                with gr.Accordion("Advanced Options", open=False):
                    filter_radius0.render()
                    resample_sr0.render()
                    rms_mix_rate0.render()
                    protect0.render() 

        vc_split.click(
            fn=cut_vocal_and_inst, 
            inputs=[vc_link, vc_download_audio, vc_split_model], 
            outputs=[vc_vocal_preview, vc_inst_preview, vc_audio_preview, vc_input]
        )
        
        authStr=", ".join(authors)
        gr.Markdown(
            "## <center>Credit to:\n"
            "#### <center>Original devs:\n"
            "<center>the RVC Project, lj1995, zomehwh, sysf\n\n"
            "#### <center>Model creators:\n"
            f"<center>{authStr}\n"
        )

if limitation is True:
    app.queue(max_size=20, api_open=config.api).launch(allowed_paths=["/"])
else:
    app.queue(max_size=20, api_open=config.api).launch(allowed_paths=["/"], share=False)