app.py

import torch
import torchaudio
from sgmse.model import ScoreModel
import gradio as gr
from sgmse.util.other import pad_spec

# Define the necessary arguments
class Args:
    device = 'cpu'  # or 'cuda' if GPU is available and enabled in the environment
    corrector = 'langevin'  # Define your corrector method
    N = 50  # Example value for number of steps
    corrector_steps = 1  # Number of steps for the corrector
    snr = 0.1  # Signal-to-noise ratio value for the corrector
    pad_mode = 'reflect'  # Pad mode for spectrogram padding

args = Args()

# Load the pre-trained model
model = ScoreModel.load_from_checkpoint("https://huggingface.co/sp-uhh/speech-enhancement-sgmse/resolve/main/train_vb_29nqe0uh_epoch%3D115.ckpt")

def enhance_speech(audio_file):
    # Load and process the audio file
    y, sr = torchaudio.load(audio_file)
    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)
    Y = pad_spec(Y, mode=args.pad_mode)
    
    # 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().numpy(), sr)
    
    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()