cleanunet is a Python package that provides an easy-to-use interface for speech denoising, based on the official PyTorch implementation of CleanUNet. This package allows users to perform high-quality speech denoising with minimal setup, leveraging the power of CleanUNet directly from PyPI.
CleanUNet is a state-of-the-art model for causal speech denoising on raw waveforms, utilizing an encoder-decoder architecture with self-attention blocks. It was originally developed by Kong, Zhifeng; Ping, Wei; Dantrey, Ambrish; and Catanzaro, Bryan. Full details can be found in their paper: Speech Denoising in the Waveform Domain with Self-Attention.
To install cleanunet, simply use pip:
pip install cleanunetcleanunet can be used to denoise audio files easily. Here's a basic example:
import time
import torch
import torchaudio
from cleanunet import CleanUNet
input_filename = "path_to_your_audio_file.wav"
output_filename = "path_to_the_denoised_audio_file.wav"
model_variant = "full" # or "high"
print(f"Loading {input_filename}...")
aud, sr = torchaudio.load(input_filename)
print(f"Loading model variant {model_variant}...")
net = CleanUNet.from_pretrained(model_variant)
print("Denoising...")
start = time.time()
with torch.no_grad():
denoised_aud = net(aud)[0]
end = time.time()
duration = aud.shape[1] / sr
print(f"Inference took {end - start:.3f}s (audio duration: {duration:.3f}s, {duration / (end - start):.1f}x)")
print(f"Saving result to {output_filename}...")
torchaudio.save(output_filename, denoised_aud, sr)To utilize CUDA for faster processing:
aud, sr = torchaudio.load("path_to_your_audio_file.wav")
net = CleanUNet.from_pretrained(variant='full', device='cuda')
with torch.no_grad():
denoised_aud = net(aud.to('cuda'))[0]This package is a pip-installable version of the CleanUNet model, as described in the paper by Kong et al. The original implementation and more comprehensive details can be found in their GitHub repository.
The structure and distributed training are adapted from WaveGlow (PyTorch), and other components are adapted from various sources as detailed in the original CleanUNet repository.
If you use cleanunet or CleanUNet in your research, please cite the original paper:
@inproceedings{kong2022speech,
title={Speech Denoising in the Waveform Domain with Self-Attention},
author={Kong, Zhifeng and Ping, Wei and Dantrey, Ambrish and Catanzaro, Bryan},
booktitle={ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
pages={7867--7871},
year={2022},
organization={IEEE}
}
This package is distributed under the same license as the original CleanUNet implementation. Please refer to the original repository for license details.