File size: 2,269 Bytes
740a68f 8a51868 a297d3b d0dbc2d 8a4b264 740a68f 8a4b264 8a51868 740a68f e6e1f50 8a4b264 a297d3b 8a4b264 a297d3b 8a4b264 1acbf3a 8a4b264 a297d3b 8a4b264 a297d3b 8a4b264 a297d3b 8a4b264 740a68f 8a4b264 740a68f 38a61d7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 |
import torch
import torchaudio
from sgmse.model import ScoreModel
import gradio as gr
from sgmse.util.other import pad_spec
# 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):
# 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="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)
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()
|