https://github.com/diffapf/la-2a
Feed-forward compressor experiments source code for "Differentiable All-pole Filters for Time-varying Audio Systems".
https://github.com/diffapf/la-2a
Last synced: 9 months ago
JSON representation
Feed-forward compressor experiments source code for "Differentiable All-pole Filters for Time-varying Audio Systems".
- Host: GitHub
- URL: https://github.com/diffapf/la-2a
- Owner: DiffAPF
- License: mpl-2.0
- Created: 2024-03-19T18:02:05.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2024-06-10T16:14:09.000Z (almost 2 years ago)
- Last Synced: 2025-08-29T16:08:16.997Z (9 months ago)
- Language: Python
- Homepage: https://diffapf.github.io/web/
- Size: 45.9 KB
- Stars: 21
- Watchers: 2
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
Differentiable All-pole Filters for Time-varying Audio Systems
Chin-Yun Yu,
Christopher Mitcheltree,
Alistair Carson,
Stefan Bilbao,
Joshua D. Reiss, and
György Fazekas
[](https://arxiv.org/abs/2404.07970)
[](https://diffapf.github.io/web/)
[](https://diffapf.github.io/web/index.html#plugins)
[](https://www.mozilla.org/en-US/MPL/2.0/FAQ/)
Feed-forward Compressor (LA-2A) Experiments
## Getting started
First, please install the required packages, including our differentiable compressor [torchcomp](https://github.com/yoyololicon/torchcomp), by running:
```bash
pip install -r requirements.txt
```
## Training
Firstly, you need to download the SignalTrain dataset from [here](https://zenodo.org/records/3824876).
The training configurations are listed under `cfg/`.
Each configurations listed under `cfg/data` corresponds to a dataset.
Please modify the `input` and `target` path of `cfg/data/la2a*.yaml` to the files of the dataset you downloaded.
To train the proposed differentiable feed-forward compressor, run:
```bash
python train_comp.py data=la2a_50
```
The training logs will be uploaded to your wandb account under the project `dafx24`.
In this example, the model is trained with peak reduction of 50.
Change the `data` argument to `la2a_75` or `la2a_25` to train the model with peak reduction of 75 or 25, respectively.
To train the frequency-sampling compressor (similar to [DASP](https://github.com/csteinmetz1/dasp-pytorch)), run:
```bash
python train_comp.py data=la2a_50 compressor.simple=true compressor.freq_sampling=true
```
A `ckpt.yaml` will be created under the logging folder (under `outputs/` by default) after training, which contains the parameters of the lowest training loss model.
We also provide our trained parameters under the folder `learned_params/`, with filenames as `[method]_[peak_reduction].yaml`.
## Evaluation
You can use your checkpoints `ckpt.yaml` or our provided learned parameters to evaluate the compressor.
Given a wave file, you can compress it using the following command:
```bash
python test_comp.py ckpt.yaml input.wav output.wav
```
## Additional notes
- `cfg/data/ff_*.yaml` are configurations for the feed-forward compressor experiments (FF-A/B/C in the paper). Please use `digital_compressor.py` to get the targets if you want to reproduce the experiments.
## Links
- [torchcomp](https://github.com/yoyololicon/torchcomp): Differentiable compressor implementation.
- [training logs](https://wandb.ai/iamycy/torchcomp-la2a/): All training logs of the compressor experiments in the paper.
## Citation
```bibtex
@inproceedings{ycy2024diffapf,
title={Differentiable All-pole Filters for Time-varying Audio Systems},
author={Chin-Yun Yu and Christopher Mitcheltree and Alistair Carson and Stefan Bilbao and Joshua D. Reiss and György Fazekas},
booktitle={International Conference on Digital Audio Effects (DAFx)},
year={2024}
}
```