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

Technical Summary: AudioSAE

Problem Statement
While Sparse Autoencoders (SAEs) have become a standard tool for interpreting dense neural representations in Natural Language Processing (NLP) and Computer Vision, their application to audio-processing models remains significantly underexplored. Existing audio models, such as Whisper and HuBERT, learn complex representations from raw audio for tasks like speech recognition and general audio understanding. However, the internal mechanisms of these models—specifically how they encode acoustic, semantic, and paralinguistic information—lack systematic interpretability. Current methods often rely on neuron-level analysis or embedding clustering, which may not effectively disentangle the polysemantic features inherent in transformer layers. There is a need to determine if SAEs can successfully decompose audio representations into sparse, interpretable, and actionable components.

Methodology
The authors train Sparse Autoencoders on the encoder activations of two prominent audio models: Whisper (an encoder-decoder Transformer trained on 680k hours of multilingual speech) and HuBERT (an encoder-only Transformer trained on 60k hours of English speech).

• Architecture & Training: The SAEs utilize a Batch-Top-k non-linearity, which was empirically selected for its superior balance of reconstruction quality and sparsity compared to Jump-ReLU and Top-k. The models are trained using an L2 reconstruction loss without auxiliary regularization on a diverse corpus of ~2.8k hours of audio (speech, music, environmental sounds).
• Evaluation Framework: The study employs a multi-faceted evaluation strategy adapted from NLP benchmarks (Karvonen et al., 2025):
  1. Stability: A distributional similarity metric (Intersection-over-Union of binary activation patterns) measures feature consistency across random seeds, layers, and model architectures.
  2. Domain Specialization: Features are categorized by activation frequency across speech, music, and environmental sound domains at both frame and audio levels.
  3. Interpretability: The authors utilize top-k probing, unlearning (feature removal), manual inspection of top-activated samples, phoneme alignment, and auto-interpretation via audio captioning and Large Language Models (LLMs).
  4. Practical Utility: The SAEs are used to steer the Whisper model to reduce hallucinations (false speech detection in non-speech segments).
  5. Neuroscientific Alignment: SAE features are correlated with human EEG activity during speech perception using Temporal Response Functions (TRFs).

Key Contributions

1. First Large-Scale Audio SAE Analysis: The paper provides the first systematic training and analysis of SAEs across all encoder layers of Whisper and HuBERT, releasing the models and code.
2. Validation of Stability and Disentanglement: The authors demonstrate that over 50% of SAE features remain consistent across different random seeds. They show that these features effectively disentangle general acoustic/semantic information from specific events (e.g., laughter, whispering, sneezing).
3. Quantitative Unlearning: The study quantifies the "cost" of erasing concepts in audio models. Removing specific speech concepts (e.g., the vowel "A") requires suppressing 19–27% of features, indicating a higher redundancy and distributed nature of phonetic information compared to text-based SAEs.
4. Practical Steering: The authors successfully apply SAE features to reduce Whisper's false speech detections (hallucinations) by 70% (reducing False Positive Rate from 0.37 to 0.11) with negligible impact on Word Error Rate (WER).
5. Neural Correlation: The work establishes a statistically significant correlation between specific SAE features and human EEG activity (specifically the Pz electrode) during speech perception, suggesting alignment between model representations and human neural processing.

Results

• Feature Consistency: Feature coverage between checkpoints trained with different seeds exceeds 50%, confirming robustness. However, features learned by HuBERT do not align with those from Whisper, likely due to differences in pre-training objectives and data.
• Layer Specialization: Whisper exhibits pronounced specialization for music features in mid-layers (peaking around layer 5), while speech-related features show distinct patterns at both frame and audio levels.
• Semantic Capture: SAE features successfully encode phonemes (achieving ~0.9 accuracy in phoneme alignment) and paralinguistic sounds. Auto-interpretation via captioning identified unique acoustic events like "ringing alarms" and "birds chirping," though it struggled with fine-grained phonetic details due to the limitations of the underlying captioning model.
• Hallucination Reduction: Steering the top-100 SAE features in Whisper reduced the False Positive Rate (FPR) by 70% across non-speech datasets (Musan, FSD50k, WHAM) while maintaining a WER increase of less than 0.4% (5.1% to 5.5%).
• EEG Correlation: Specific SAE features showed significant correlation with EEG signals at time lags between 0 and 500ms, particularly for features activating on specific vowels.

Significance and Claims
The paper claims to fill a critical gap in audio model interpretability by demonstrating that SAEs are effective tools for decomposing dense audio representations. The authors assert that their work proves SAE features serve dual roles: as interpretable units that reveal how models encode acoustic and semantic information, and as actionable levers for steering model behavior (e.g., reducing hallucinations). Furthermore, the correlation with EEG activity suggests that these learned features align with human neural processing, providing a bridge between artificial and biological audio perception. The authors maintain a modest stance, acknowledging that their evaluation is limited to specific classification tasks and base/small model variants, and that future work is needed to explore broader audio tasks and more sophisticated non-linear brain models.