https://github.com/haoheliu/torchsubband
Pytorch implementation of subband decomposition
https://github.com/haoheliu/torchsubband
deep-learning music-source-separation signal-processing speech-enhancement speech-processing speech-recognition
Last synced: 3 months ago
JSON representation
Pytorch implementation of subband decomposition
- Host: GitHub
- URL: https://github.com/haoheliu/torchsubband
- Owner: haoheliu
- License: mit
- Created: 2021-09-19T08:09:49.000Z (over 3 years ago)
- Default Branch: main
- Last Pushed: 2022-07-26T11:26:27.000Z (almost 3 years ago)
- Last Synced: 2025-02-21T03:47:56.393Z (3 months ago)
- Topics: deep-learning, music-source-separation, signal-processing, speech-enhancement, speech-processing, speech-recognition
- Language: HTML
- Homepage: https://haoheliu.github.io/torchsubband/
- Size: 374 KB
- Stars: 92
- Watchers: 3
- Forks: 13
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
[](https://badge.fury.io/py/torchsubband)
# torchsubband
This's a package for subband decomposition.
It can transform waveform into three kinds of revertable subband feature representations, which are potentially useful features for music source separation or similar tasks.
[](https://imgtu.com/i/488zR0)
## Usage
Installation
```shell
pip install torchsubband
```
A simple example:
```python
from torchsubband import SubbandDSP
import torch
# nn.Module
model = SubbandDSP(subband=2) # You can choose 1,2,4, or 8
batchsize=3 # any int number
channel=1 # any int number
length = 44100*2 # any int number
input = torch.randn((batchsize,channel,length))
# Get subband waveform
subwav = model.wav_to_sub(input)
reconstruct_1 = model.sub_to_wav(subwav,length=length)
# Get subband magnitude spectrogram
sub_spec,cos,sin = model.wav_to_mag_phase_sub_spec(input)
reconstruct_2 = model.mag_phase_sub_spec_to_wav(sub_spec,cos,sin,length=length)
# Get subband complex spectrogram
sub_complex_spec = model.wav_to_complex_sub_spec(input)
reconstruct_3 = model.complex_sub_spec_to_wav(sub_complex_spec,length=length)
```
## Reconstruction loss
The following table shows the reconstruction quality. We tried a set of audio to conduct subband decomposition and reconstruction.
| Subbands | L1loss | PESQ | SiSDR|
| :----: | :----: | :----: | :----:
| 2 | 1e-6 | 4.64 | 61.8 |
| 4 | 1e-6 | 4.64 | 58.9 |
| 8 | 5e-5 | 4.64 | 58.2 |
You can also test this program by running the following test script. It will give you some evaluation output.
```python
from torchsubband import test
test()
```
## Citation
If you find our code useful for your research, please consider citing:
```bib
@misc{liu2021cwspresunet,
title={CWS-PResUNet: Music Source Separation with Channel-wise Subband Phase-aware ResUNet},
author={Haohe Liu and Qiuqiang Kong and Jiafeng Liu},
year={2021},
eprint={2112.04685},
archivePrefix={arXiv},
primaryClass={cs.SD}
}
```
```bib
@inproceedings{Liu2020,
author={Haohe Liu and Lei Xie and Jian Wu and Geng Yang},
title={{Channel-Wise Subband Input for Better Voice and Accompaniment Separation on High Resolution Music}},
year=2020,
booktitle={Proc. Interspeech 2020},
pages={1241--1245},
doi={10.21437/Interspeech.2020-2555},
url={http://dx.doi.org/10.21437/Interspeech.2020-2555}
}
```