AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders

TL;DR

This paper explores the use of Sparse Autoencoders to interpret and manipulate audio-processing neural models, demonstrating their stability, interpretability, and practical utility in real-world applications.

Contribution

It introduces a novel application of SAEs across all encoder layers of Whisper and HuBERT, providing extensive evaluation and demonstrating their effectiveness in capturing and disentangling audio features.

Findings

Over 50% feature consistency across seeds

SAE features capture both acoustic and semantic info

Feature steering reduces false speech detections by 70%

Abstract

Sparse Autoencoders (SAEs) are powerful tools for interpreting neural representations, yet their use in audio remains underexplored. We train SAEs across all encoder layers of Whisper and HuBERT, provide an extensive evaluation of their stability, interpretability, and show their practical utility. Over 50% of the features remain consistent across random seeds, and reconstruction quality is preserved. SAE features capture general acoustic and semantic information as well as specific events, including environmental noises and paralinguistic sounds (e.g. laughter, whispering) and disentangle them effectively, requiring removal of only 19-27% of features to erase a concept. Feature steering reduces Whisper's false speech detections by 70% with negligible WER increase, demonstrating real-world applicability. Finally, we find SAE features correlated with human EEG activity during speech perception, indicating alignment with human neural processing. The code and checkpoints are available at https://github.com/audiosae/audiosae_demo.