app.py

import torch
import torchaudio
from sgmse.model import ScoreModel
import gradio as gr
from sgmse.util.other import pad_spec
import time  # Import the time module

# Define parameters based on the argparse configuration in enhancement.py
args = {
    "test_dir": "./test_data",  # example directory, adjust as needed
    "enhanced_dir": "./enhanced_data",  # example directory, adjust as needed
    "ckpt": "https://huggingface.co/sp-uhh/speech-enhancement-sgmse/resolve/main/train_vb_29nqe0uh_epoch%3D115.ckpt",
    "corrector": "ald",
    "corrector_steps": 1,
    "snr": 0.5,
    "N": 30,
    "device": "cuda" if torch.cuda.is_available() else "cpu"
}

# Load the pre-trained model
model = ScoreModel.load_from_checkpoint(args["ckpt"])

def enhance_speech(audio_file):
    start_time = time.time()  # Start the timer

    # Load and process the audio file
    y, sr = torchaudio.load(audio_file)
    print(f"Loaded audio in {time.time() - start_time:.2f}s")
    T_orig = y.size(1)

    # Normalize
    norm_factor = y.abs().max()
    y = y / norm_factor

    # Prepare DNN input
    Y = torch.unsqueeze(model._forward_transform(model._stft(y.to(args["device"]))), 0)
    print(f"Transformed input in {time.time() - start_time:.2f}s")

    Y = pad_spec(Y, mode="zero_pad")  # Use "zero_pad" mode for padding

    # Reverse sampling
    sampler = model.get_pc_sampler(
        'reverse_diffusion', args["corrector"], Y.to(args["device"]),
        N=args["N"], corrector_steps=args["corrector_steps"], snr=args["snr"]
    )
    sample, _ = sampler()

    # Backward transform in time domain
    x_hat = model.to_audio(sample.squeeze(), T_orig)

    # Renormalize
    x_hat = x_hat * norm_factor

    # Save the enhanced audio
    output_file = 'enhanced_output.wav'
    torchaudio.save(output_file, x_hat.cpu(), sr)

    print(f"Processed audio in {time.time() - start_time:.2f}s")
    return output_file

# Gradio interface setup
inputs = gr.Audio(label="Input Audio", type="filepath")
outputs = gr.Audio(label="Output Audio", type="filepath")
title = "Speech Enhancement using SGMSE"
description = "This Gradio demo uses the SGMSE model for speech enhancement. Upload your audio file to enhance it."
article = "<p style='text-align: center'><a href='https://huggingface.co/SP-UHH/speech-enhancement-sgmse' target='_blank'>Model Card</a></p>"

# Launch without share=True (as it's not supported on Hugging Face Spaces)
gr.Interface(fn=enhance_speech, inputs=inputs, outputs=outputs, title=title, description=description, article=article).launch()