app.py

import os
import gradio as gr
from pydub import AudioSegment
from audio_separator.separator import Separator
from lib.infer import infer_audio

# Define a function to handle the entire separation process
def separate_audio(input_audio, output_dir, model_voc_inst, model_deecho, model_back_voc):
    # Create output directory if it doesn't exist
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)

    separator = Separator(output_dir=output_dir)

    # Define output files
    vocals = os.path.join(output_dir, 'Vocals.wav')
    instrumental = os.path.join(output_dir, 'Instrumental.wav')
    vocals_reverb = os.path.join(output_dir, 'Vocals (Reverb).wav')
    vocals_no_reverb = os.path.join(output_dir, 'Vocals (No Reverb).wav')
    lead_vocals = os.path.join(output_dir, 'Lead Vocals.wav')
    backing_vocals = os.path.join(output_dir, 'Backing Vocals.wav')

    # Splitting a track into Vocal and Instrumental
    separator.load_model(model_filename=model_voc_inst)
    voc_inst = separator.separate(input_audio)
    os.rename(os.path.join(output_dir, voc_inst[0]), instrumental)  # Rename to “Instrumental.wav”
    os.rename(os.path.join(output_dir, voc_inst[1]), vocals)        # Rename to “Vocals.wav”

    # Applying DeEcho-DeReverb to Vocals
    separator.load_model(model_filename=model_deecho)
    voc_no_reverb = separator.separate(vocals)
    os.rename(os.path.join(output_dir, voc_no_reverb[0]), vocals_no_reverb)  # Rename to “Vocals (No Reverb).wav”
    os.rename(os.path.join(output_dir, voc_no_reverb[1]), vocals_reverb)     # Rename to “Vocals (Reverb).wav”

    # Separating Back Vocals from Main Vocals
    separator.load_model(model_filename=model_back_voc)
    backing_voc = separator.separate(vocals_no_reverb)
    os.rename(os.path.join(output_dir, backing_voc[0]), backing_vocals)  # Rename to “Backing Vocals.wav”
    os.rename(os.path.join(output_dir, backing_voc[1]), lead_vocals)     # Rename to “Lead Vocals.wav”

    return instrumental, vocals, vocals_reverb, vocals_no_reverb, lead_vocals, backing_vocals


# Main function to process audio (Inference)
def process_audio(MODEL_NAME, SOUND_PATH, F0_CHANGE, F0_METHOD, MIN_PITCH, MAX_PITCH, CREPE_HOP_LENGTH, INDEX_RATE, 
                  FILTER_RADIUS, RMS_MIX_RATE, PROTECT, SPLIT_INFER, MIN_SILENCE, SILENCE_THRESHOLD, SEEK_STEP, 
                  KEEP_SILENCE, FORMANT_SHIFT, QUEFRENCY, TIMBRE, F0_AUTOTUNE, OUTPUT_FORMAT, upload_audio=None):

    # If no sound path is given, use the uploaded file
    if not SOUND_PATH and upload_audio is not None:
        SOUND_PATH = os.path.join("uploaded_audio", upload_audio.name)
        with open(SOUND_PATH, "wb") as f:
            f.write(upload_audio.read())
    
    # Check if a model name is provided
    if not MODEL_NAME:
        return "Please provide a model name."

    # Run the inference
    os.system("chmod +x stftpitchshift")
    inferred_audio = infer_audio(
        MODEL_NAME,
        SOUND_PATH,
        F0_CHANGE,
        F0_METHOD,
        MIN_PITCH,
        MAX_PITCH,
        CREPE_HOP_LENGTH,
        INDEX_RATE,
        FILTER_RADIUS,
        RMS_MIX_RATE,
        PROTECT,
        SPLIT_INFER,
        MIN_SILENCE,
        SILENCE_THRESHOLD,
        SEEK_STEP,
        KEEP_SILENCE,
        FORMANT_SHIFT,
        QUEFRENCY,
        TIMBRE,
        F0_AUTOTUNE,
        OUTPUT_FORMAT
    )
    
    return inferred_audio


# Gradio Blocks Interface with Tabs
with gr.Blocks(title="Hex RVC") as app:
    gr.Markdown("# Hex RVC")
    
    with gr.Tab("Audio Separation"):
        with gr.Row():
            input_audio = gr.Audio(source="upload", type="filepath", label="Upload Audio File")
            output_dir = gr.Textbox(value="/content/output", label="Output Directory")
        
        with gr.Row():
            model_voc_inst = gr.Textbox(value='model_bs_roformer_ep_317_sdr_12.9755.ckpt', label="Vocal & Instrumental Model")
            model_deecho = gr.Textbox(value='UVR-DeEcho-DeReverb.pth', label="DeEcho-DeReverb Model")
            model_back_voc = gr.Textbox(value='mel_band_roformer_karaoke_aufr33_viperx_sdr_10.1956.ckpt', label="Backing Vocals Model")
        
        separate_button = gr.Button("Separate Audio")
        
        with gr.Row():
            instrumental_out = gr.Audio(label="Instrumental")
            vocals_out = gr.Audio(label="Vocals")
            vocals_reverb_out = gr.Audio(label="Vocals (Reverb)")
            vocals_no_reverb_out = gr.Audio(label="Vocals (No Reverb)")
            lead_vocals_out = gr.Audio(label="Lead Vocals")
            backing_vocals_out = gr.Audio(label="Backing Vocals")
        
        separate_button.click(
            separate_audio,
            inputs=[input_audio, output_dir, model_voc_inst, model_deecho, model_back_voc],
            outputs=[instrumental_out, vocals_out, vocals_reverb_out, vocals_no_reverb_out, lead_vocals_out, backing_vocals_out]
        )
    
    with gr.Tab("Inference"):
        with gr.Row():
            MODEL_NAME = gr.Textbox(label="Model Name", placeholder="Enter model name")
            SOUND_PATH = gr.Textbox(label="Audio Path (Optional)", placeholder="Leave blank to upload audio")
            upload_audio = gr.File(label="Upload Audio", type='filepath', file_types=["audio"])
        
        with gr.Row():
            F0_CHANGE = gr.Number(label="Pitch Change (semitones)", value=0)
            F0_METHOD = gr.Dropdown(choices=["crepe", "harvest", "mangio-crepe", "rmvpe", "rmvpe+", "fcpe", 
                                             "hybrid[mangio-crepe+rmvpe]", "hybrid[mangio-crepe+fcpe]", 
                                             "hybrid[rmvpe+fcpe]", "hybrid[mangio-crepe+rmvpe+fcpe]"], 
                                    label="F0 Method", value="fcpe")
        
        with gr.Row():
            MIN_PITCH = gr.Textbox(label="Min Pitch", value="50")
            MAX_PITCH = gr.Textbox(label="Max Pitch", value="1100")
            CREPE_HOP_LENGTH = gr.Number(label="Crepe Hop Length", value=120)
            INDEX_RATE = gr.Slider(label="Index Rate", minimum=0, maximum=1, value=0.75)
            FILTER_RADIUS = gr.Number(label="Filter Radius", value=3)
            RMS_MIX_RATE = gr.Slider(label="RMS Mix Rate", minimum=0, maximum=1, value=0.25)
            PROTECT = gr.Slider(label="Protect", minimum=0, maximum=1, value=0.33)
        
        with gr.Accordion("Advanced Settings", open=False):
            SPLIT_INFER = gr.Checkbox(label="Enable Split Inference", value=False)
            MIN_SILENCE = gr.Number(label="Min Silence (ms)", value=500)
            SILENCE_THRESHOLD = gr.Number(label="Silence Threshold (dBFS)", value=-50)
            SEEK_STEP = gr.Slider(label="Seek Step (ms)", minimum=1, maximum=10, value=1)
            KEEP_SILENCE = gr.Number(label="Keep Silence (ms)", value=200)
            FORMANT_SHIFT = gr.Checkbox(label="Enable Formant Shift", value=False)
            QUEFRENCY = gr.Number(label="Quefrency", value=0)
            TIMBRE = gr.Number(label="Timbre", value=1)
            F0_AUTOTUNE = gr.Checkbox(label="Enable F0 Autotune", value=False)
            OUTPUT_FORMAT = gr.Dropdown(choices=["wav", "flac", "mp3"], label="Output Format", value="wav")
        
        run_button = gr.Button("Run Inference")
        output_audio = gr.Audio(label="Generated Audio", type='filepath')

        run_button.click(
            process_audio, 
            inputs=[MODEL_NAME, SOUND_PATH, F0_CHANGE, F0_METHOD, MIN_PITCH, MAX_PITCH, CREPE_HOP_LENGTH, INDEX_RATE, 
                    FILTER_RADIUS, RMS_MIX_RATE, PROTECT, SPLIT_INFER, MIN_SILENCE, SILENCE_THRESHOLD, SEEK_STEP, 
                    KEEP_SILENCE, FORMANT_SHIFT, QUEFRENCY, TIMBRE, F0_AUTOTUNE, OUTPUT_FORMAT, upload_audio], 
            outputs=output_audio
        )

# Launch the Gradio app
app.launch()