app.py

import torch
import librosa
import soundfile as sf
import numpy as np
import tempfile
import gradio as gr

from denoiser import pretrained
from denoiser.dsp import convert_audio
from pydub import AudioSegment, silence
from tqdm import tqdm


# -----------------------------
# Load model ONCE
# -----------------------------
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = pretrained.dns64().to(device)


# -----------------------------
# Silence trimming helpers
# -----------------------------
def safe_append(base, chunk, crossfade_ms=30):
    if len(base) > 0 and len(chunk) > 0:
        safe_crossfade = min(crossfade_ms, len(base), len(chunk))
        if safe_crossfade > 0:
            return base.append(chunk, crossfade=safe_crossfade)
    return base + chunk


def shorten_silences(audio, silence_thresh=-50, min_silence_len=400, max_keep_silence=1500, crossfade_ms=30):
    """
    Detects silences and reduces them using a stepped approach.
    """
    silent_ranges = silence.detect_silence(
        audio,
        min_silence_len=min_silence_len,
        silence_thresh=silence_thresh
    )

    output = AudioSegment.silent(duration=0)
    prev_end = 0

    for start, end in silent_ranges:
        # Add the non-silent chunk
        chunk = audio[prev_end:start]
        output = safe_append(output, chunk, crossfade_ms)

        silence_len = end - start
        
        # Smart stepped reduction logic:
        # - Short pauses (< 500ms) are kept as is.
        # - Longer ones are compressed in steps of 500ms for every 1000ms extra.
        if silence_len < 500:
            keep = silence_len
        else:
            # Calculation: starts at 500ms, adds 500ms for every full second beyond the first.
            keep = min(max_keep_silence, (silence_len // 1000) * 500 + 500)

        output = safe_append(
            output,
            AudioSegment.silent(duration=int(keep)),
            crossfade_ms
        )
        prev_end = end

    # Add final chunk
    output = safe_append(output, audio[prev_end:], crossfade_ms)
    return output


# -----------------------------
# Main processing function
# -----------------------------
def denoise_audio(audio_file, trim_silence, silence_thresh, min_silence_len, max_keep_silence):
    if audio_file is None:
        return None
        
    # Load audio
    wav, sr = librosa.load(audio_file, sr=16000)

    chunk_size = 16000 * 10
    denoised_chunks = []

    # Process in chunks to avoid OOM
    for i in range(0, len(wav), chunk_size):
        chunk = wav[i:i + chunk_size]
        wav_tensor = torch.tensor(chunk).unsqueeze(0).to(device)
        wav_tensor = convert_audio(
            wav_tensor, sr, model.sample_rate, model.chin
        )

        with torch.no_grad():
            denoised = model(wav_tensor)[0]

        denoised_chunks.append(
            denoised.squeeze().cpu().numpy()
        )

    denoised_np = np.concatenate(denoised_chunks)

    # Save denoised to temp
    tmp_wav = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
    sf.write(tmp_wav.name, denoised_np, model.sample_rate)

    if trim_silence:
        audio = AudioSegment.from_file(tmp_wav.name, format="wav")
        processed = shorten_silences(
            audio, 
            silence_thresh=silence_thresh, 
            min_silence_len=min_silence_len,
            max_keep_silence=max_keep_silence
        )
        final_file = tempfile.NamedTemporaryFile(
            suffix="_final.wav", delete=False
        )
        processed.export(final_file.name, format="wav")
        return final_file.name

    return tmp_wav.name


# -----------------------------
# Gradio UI
# -----------------------------
with gr.Blocks(title="🎧 Advanced Audio Denoiser") as demo:
    gr.Markdown("# 🎧 Audio Denoiser (Demucs DNS64)")
    gr.Markdown("Upload audio to remove noise and intelligently shorten silences.")
    
    with gr.Row():
        with gr.Column():
            input_audio = gr.Audio(type="filepath", label="Upload Audio")
            
            with gr.Accordion("Silence Trimming Settings", open=True):
                do_trim = gr.Checkbox(label="Enable Silence Trimming", value=True)
                
                threshold = gr.Slider(
                    minimum=-70, maximum=-20, value=-50, step=1,
                    label="Silence Threshold (dB)",
                    info="Lower is stricter (quieter sounds count as silence)"
                )
                
                min_len = gr.Slider(
                    minimum=100, maximum=2000, value=400, step=50,
                    label="Min Silence Duration (ms)",
                    info="How long a pause must be to be considered 'silence'"
                )
                
                max_keep = gr.Slider(
                    minimum=0, maximum=5000, value=1500, step=100,
                    label="Max Silence to Keep (ms)",
                    info="Longer silences will be shortened towards this value"
                )
            
            submit_btn = gr.Button("Process Audio", variant="primary")
            
        with gr.Column():
            output_audio = gr.Audio(label="Denoised & Trimmed Output")

    submit_btn.click(
        fn=denoise_audio,
        inputs=[input_audio, do_trim, threshold, min_len, max_keep],
        outputs=output_audio
    )

if __name__ == "__main__":
    demo.launch()