app.py

from audiotools import AudioSignal
from pyharp.core import ModelCard, build_endpoint
from pyharp.media.audio import load_audio, save_audio
from pyharp.tags import Subcategory

import gradio as gr
import librosa
import torch

from pathlib import Path
import time

# ModelCard

model_card = ModelCard(
    name="Harmonic-Percussive Source Separation",
    description=(
        "Splits a track into its harmonic and percussive components and remixes them at an "
        "adjustable balance, using librosa's harmonic-percussive source separation. Repository: "
        "https://github.com/librosa/librosa"
    ),
    author="Brian McFee, librosa developers",
    tags=[Subcategory.STEM_SEPARATION, 'example']
)


# Model Logic
def hpss(signal: AudioSignal, **kwargs):
    h, p = librosa.effects.hpss(signal.audio_data.squeeze().numpy(), **kwargs)

    if h.ndim == 1:
        h = h[None, None, :]
        p = p[None, None, :]
    elif h.ndim == 2:
        h = h[None, :, :]
        p = p[None, :, :]
    else:
        raise ValueError("Unexpected audio shape")

    harmonic_signal = signal.clone()
    harmonic_signal.audio_data = torch.from_numpy(h)

    percussive_signal = signal.clone()
    percussive_signal.audio_data = torch.from_numpy(p)

    return harmonic_signal, percussive_signal

MIN_DB = -120

def process_fn(audio_file_path,
               harmonic_db: float, 
               percussive_db: float, 
               kernel_size: int = 31, 
               margin: float = 1.0):
    sig = load_audio(audio_file_path)
    
    harmonic, percussive = hpss(sig, kernel_size=int(kernel_size), margin=margin)

    def clip(db):
        return -float("inf") if db == MIN_DB else db

    # Mix signals with gain
    output_sig = (
        harmonic.volume_change(clip(harmonic_db)) +
        percussive.volume_change(clip(percussive_db))
    )

    #filename
    input_path = Path(audio_file_path)
    stem = input_path.stem 
    filename = f"{stem}_hpss.wav"
    output_audio_path = save_audio(output_sig, filename)
    return output_audio_path


# Gradio Components
input_audio = gr.Audio(type="filepath", label="Input Audio").harp_required(True).set_info((
    "The audio to split. Harmonic content is the sustained, pitched "
    "part and percussive content is the transient part."
))

harmonic_slider = gr.Slider(
    minimum=MIN_DB, maximum=24, 
    step=1, value=0, 
    label="Harmonic Level (dB)",
    info="Boost or reduce tonal parts like chords and pads."
)

percussive_slider = gr.Slider(
    minimum=MIN_DB, maximum=24, 
    step=1, value=0, 
    label="Percussive Level (dB)",
    info="Control the loudness of drums and other rhythmic hits."
)

kernel_slider = gr.Slider(
    minimum=1, maximum=101, 
    step=1, value=31, 
    label="Time Resolution",
    info="How much time/spectral context is used to classify energy as harmonic or percussive. \n Lower = more detailed, reacts faster to rhythm, higher = smoother and slower"
)

margin_slider = gr.Slider(
    minimum=0.5, maximum=5.0, 
    step=0.1, value=1.0, 
    label="Separation Strength",
    info="How strongly harmonic & percussive components must differ to be separated. \n Higher = cleaner isolation, lower = more natural blend."
)

output_audio = gr.Audio(type="filepath", label="Remixed Audio").set_info((
    "The harmonic and percussive parts, remixed by the balance control."
))


# Launch the App 
if __name__ == "__main__":
    with gr.Blocks() as demo:
        # Build endpoint inside Blocks context
        endpoint = build_endpoint(
            model_card=model_card,
            input_components=[
                input_audio,
                harmonic_slider,
                percussive_slider,
                kernel_slider,
                margin_slider
            ],
            output_components=[output_audio],
            process_fn=process_fn
        )

        # Layout
        endpoint["controls_button"]
        endpoint["controls_data"]
        input_audio.render()
        harmonic_slider.render()
        percussive_slider.render()
        kernel_slider.render()
        margin_slider.render()
        output_audio.render()
        endpoint["process_button"]
        endpoint["cancel_button"]

    demo.queue()
    demo.launch(pwa=True, show_error=True)