AudioSAE: Towards Understanding of Audio-Processing Models with Sparse AutoEncoders
This work addresses the need for better interpretability in audio AI models, offering practical tools for researchers and developers, though it is incremental as it extends existing SAE methods to the audio domain.
The paper tackled the problem of interpreting audio-processing models like Whisper and HuBERT by applying Sparse Autoencoders (SAEs) to their encoder layers, achieving results such as over 50% feature consistency across seeds, 19-27% feature removal to erase concepts, and a 70% reduction in false speech detections with minimal WER increase.
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.