Sparse Autoencoders make Audio Foundation Models more Explainable

Summary

This repository contains the code used to perform experiments presented in the paper "" accepted at ICASSP 2026 (https://arxiv.org/pdf/2509.24793).

Installation (WIP)

Clone the repository:

git clone git@github.com:theomariotte/sae_audio_ssl.git

Install the required packages (TODO):

pip install -r requirements.txt

Install the module:

cd sae_audio_ssl
pip install .

Examples (TODO)

1. Training an SAE in a given model

2. Training a probe in a given layer of a pretrained model*

3. Adding a new pretrained model from huggingface

4. Evaluating disentanglement from acoustic features

Citation

@inproceeding{mariotte2025sparseautoencodersmakeaudio,
      title={Sparse Autoencoders Make Audio Foundation Models more Explainable}, 
      author={Théo Mariotte and Martin Lebourdais and Antonio Almudévar and Marie Tahon and Alfonso Ortega and Nicolas Dugué},
      year={2025},
      eprint={2509.24793},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2509.24793}, 
}