Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Georgii Aparin, Tasnima Sadekova, Alexey Rukhovich, Assel Yermekova, Laida Kushnareva, Vadim Popov, Kristian Kuznetsov, Irina Piontkovskaya

TL;DR

This work investigates the interpretability of large audio models by training Sparse Autoencoders on activations from Whisper and HuBERT across all encoder layers. It demonstrates robust, semantically and paralinguistically rich SAE features that disentangle concepts like phonemes and environmental sounds, and shows practical benefits such as a 70% reduction in Whisper hallucinations with minimal WER impact and correlations with human EEG signals. The study introduces comprehensive evaluation metrics for SAE quality, stability, and domain specialization, and provides tools, code, and checkpoints to enable replication. Overall, SAEs offer a principled pathway to understanding and steering complex audio representations with clear, interpretable latent factors relevant to both engineering and neuroscience perspectives.

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.

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

TL;DR

This work investigates the interpretability of large audio models by training Sparse Autoencoders on activations from Whisper and HuBERT across all encoder layers. It demonstrates robust, semantically and paralinguistically rich SAE features that disentangle concepts like phonemes and environmental sounds, and shows practical benefits such as a 70% reduction in Whisper hallucinations with minimal WER impact and correlations with human EEG signals. The study introduces comprehensive evaluation metrics for SAE quality, stability, and domain specialization, and provides tools, code, and checkpoints to enable replication. Overall, SAEs offer a principled pathway to understanding and steering complex audio representations with clear, interpretable latent factors relevant to both engineering and neuroscience perspectives.

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.
Paper Structure (46 sections, 10 equations, 28 figures, 13 tables)

This paper contains 46 sections, 10 equations, 28 figures, 13 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Audio Representations
  4. SAE in Various Domains and Applications
  5. Evaluating SAE Quality
  6. Other Interpretability Methods in Audio Domain
  7. Background and Methodology
  8. SAE
  9. Audio SAE Training Setup
  10. SAE Evaluation and Analysis
  11. Feature Robustness via Distributional Semantics
  12. Domain Specialization
  13. Interpretability Analysis
  14. Hallucination Reduction via SAE Steering
  15. Correlation with EEG
  16. ...and 31 more sections

Figures (28)

  • Figure 1: The trade-off between normalized reconstruction error (L2) and sparsity (L0) for models at layer 12.
  • Figure 2: Layer-wise feature specialization ratio by speech, sounds, and music domains for Whisper (solid line) and HuBERT (dashed) at frame (top) and audio (bottom) levels.
  • Figure 3: Top-$k$ probing and unlearning for accent classification. More results in Appendix \ref{['app:classification']}.
  • Figure 4: Selective unlearning of letter 'A' via iterative feature removal. Feature indices on x-axis ordered by discriminative importance for target vowel.
  • Figure 5: Features identified via the label-based classification experiment. Each sub-figure displays a feature's values above the corresponding audio mel-spectrogram.
  • ...and 23 more figures