app.py

import os
import re
import random
from scipy.io.wavfile import write, read
import numpy as np
import gradio as gr
import yt_dlp

# Model dictionaries and lists
roformer_models = {
    'BS-Roformer-Viperx-1297.ckpt': 'model_bs_roformer_ep_317_sdr_12.9755.ckpt',
    'BS-Roformer-Viperx-1296.ckpt': 'model_bs_roformer_ep_368_sdr_12.9628.ckpt',
    'BS-Roformer-Viperx-1053.ckpt': 'model_bs_roformer_ep_937_sdr_10.5309.ckpt',
    'Mel-Roformer-Viperx-1143.ckpt': 'model_mel_band_roformer_ep_3005_sdr_11.4360.ckpt'
}

mdx23c_models = [
    'MDX23C_D1581.ckpt',
    'MDX23C-8KFFT-InstVoc_HQ.ckpt',
    'MDX23C-8KFFT-InstVoc_HQ_2.ckpt',
]

# More model lists...

output_format = ['wav', 'flac', 'mp3']
mdxnet_overlap_values = ['0.25', '0.5', '0.75', '0.99']
vrarch_window_size_values = ['320', '512', '1024']
demucs_overlap_values = ['0.25', '0.50', '0.75', '0.99']

# Function to download audio
def download_audio(url):
    ydl_opts = {
        'format': 'bestaudio/best',
        'outtmpl': 'ytdl/%(title)s.%(ext)s',
        'postprocessors': [{
            'key': 'FFmpegExtractAudio',
            'preferredcodec': 'wav',
            'preferredquality': '192',
        }],
    }
    with yt_dlp.YoutubeDL(ydl_opts) as ydl:
        info_dict = ydl.extract_info(url, download=True)
        file_path = ydl.prepare_filename(info_dict).rsplit('.', 1)[0] + '.wav'
        sample_rate, audio_data = read(file_path)
        audio_array = np.asarray(audio_data, dtype=np.int16)
        return sample_rate, audio_array

# Function to separate audio using Roformer
def roformer_separator(audio, model, output_format, overlap, segment_size, denoise):
    directory = "./outputs"
    random_id = str(random.randint(10000, 99999))
    os.makedirs("outputs", exist_ok=True)
    write(f'{random_id}.wav', audio[0], audio[1])
    full_roformer_model = roformer_models[model]
    
    prompt = f"audio-separator {random_id}.wav --model_filename {full_roformer_model} --output_dir=./outputs --output_format={output_format} --normalization=0.9 --mdxc_overlap={overlap} --mdxc_segment_size={segment_size}"
    if denoise:
        prompt += " --mdx_enable_denoise"
    
    os.system(prompt)

    files_list = [os.path.join(directory, file) for file in os.listdir(directory) if re.search(random_id, file)]
    stem1_file, stem2_file, stem3_file = files_list[:3]  # Assuming the files are in the correct order
    return stem1_file, stem2_file, stem3_file

# Gradio interface
def process_audio(url, model, output_format, overlap, segment_size, denoise):
    sample_rate, audio_array = download_audio(url)
    stems = roformer_separator((sample_rate, audio_array), model, output_format, overlap, segment_size, denoise)
    return stems

# Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# Hex Audio Separator")
    with gr.Row():
        url_input = gr.Textbox(label="YouTube URL")
        model_input = gr.Dropdown(choices=list(roformer_models.keys()), label="Roformer Model")
        format_input = gr.Dropdown(choices=output_format, label="Output Format")
        overlap_input = gr.Dropdown(choices=mdxnet_overlap_values, label="Overlap")
        segment_input = gr.Slider(0, 100, label="Segment Size")
        denoise_input = gr.Checkbox(label="Enable Denoise")

    output1 = gr.Audio(label="Vocals")
    output2 = gr.Audio(label="Instrumental")
    output3 = gr.Audio(label="Backing Vocals")

    submit_button = gr.Button("Process")
    submit_button.click(
        process_audio,
        inputs=[url_input, model_input, format_input, overlap_input, segment_input, denoise_input],
        outputs=[output1, output2, output3]
    )

demo.launch()