PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2602.05027v2cs.SDcs.AI

TL;DR

Audio-model representations remain underexplored with sparse autoencoders despite the method’s use in other domains. The paper trains and evaluates SAEs across Whisper and HuBERT, finding stable, interpretable features and practical utility for reducing Whisper hallucinations, while identifying limits in task coverage and auto-interpretation.

  • Problem

    Sparse autoencoders are powerful interpretability tools, but their application to audio representation models remains largely unexplored.

  • Method

    The paper trains sparse autoencoders on Whisper and HuBERT representations and evaluates their stability, interpretability, disentanglement, and steering utility.

  • Results

    SAE features are stable across random seeds, encode semantic, paralinguistic, and acoustic information, and reduce Whisper’s hallucination false positive rate by 70%.

  • Takeaways & Limitations

    SAEs provide interpretable audio features with practical relevance for steering Whisper and analyzing speech-related representations.

  • Takeaways & Limitations

    The downstream evaluation covers limited classification tasks and applications, while auto-interpretation can lose fine-grained phoneme information because its captioning model favors music and sound data.

Abstract

from arXiv · show

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.

1 Introduction

Audio models have become powerful and broadly useful, but their internal representations remain difficult to interpret systematically. This work applies sparse autoencoders to Whisper and HuBERT, evaluating feature stability, interpretability, and practical relevance.

  • Recent Transformer-based audio models support speech recognition, synthesis, and general audio understanding at large scale.
  • Whisper and HuBERT produce robust, semantically rich embeddings across diverse audio tasks, motivating systematic interpretability methods.Their representations are shaped by different training objectives and data.
  • Audio applications of sparse autoencoders remain largely unexplored compared with their extensive use in text and vision.Prior audio work included isolated music-modeling attempts and informal Whisper analyses.
  • The paper trains sparse autoencoders on Whisper and HuBERT activations and provides the first large-scale analysis of audio representation interpretability with this method.
  • The evaluation shows that SAE features are stable across seeds and encode semantic, paralinguistic, and acoustic information.
  • SAE features have practical and neuroscientific relevance through Whisper steering for hallucination reduction and correlations with human EEG activity.

2 Related Works

Prior work established sparse autoencoders as tools for interpretability and behavior control in language, vision, and other domains, while audio applications remained limited. This paper extends the analysis to speech-oriented models and SAE-disentangled representations, introducing broader feature analyses than earlier acoustic-neuron studies.

  • Audio Representations: Whisper and HuBERT are widely used speech-processing models with distinct architectures, training data, and objectives.Whisper is an encoder-decoder trained on 680k hours of multilingual speech, whereas HuBERT is encoder-only and trained on 60k hours of English speech.
  • Natural Language Processing: SAEs have been studied extensively in NLP for interpreting features, detecting spurious correlations, and controlling model behavior.Examples include separating gender markers from professional content and identifying features associated with harmful-prompt classification shortcuts.
  • SAE utility in downstream NLP tasks: Results for SAE utility in downstream NLP tasks are mixed, ranging from no advantage over linear probes to benefits in data selection, detection, and query steering.
  • Computer Vision: In computer vision, SAEs support model interpretability, image-generation control, and selective concept unlearning without fine-tuning.
  • Other domains: SAEs have also produced interpretable features aligned with biological annotations in protein language models.
  • Audio: Audio SAE research was previously limited, with prior work focusing on concept discovery in music samples.
  • SAE evaluation: Comprehensive SAE evaluation frameworks span reconstruction, interpretability, concept detection, and disentanglement, while showing that sparsity-fidelity improvements may not transfer to practical tasks.
  • Comparison with acoustic-neuron work: The paper complements acoustic-neuron studies by analyzing SAE-disentangled speech-model representations and identifying feature types and clusters not discussed previously.

3 Background and Methodology

This section explains how SAEs recover sparse, potentially interpretable directions from audio-model activations and how the paper evaluates their robustness, interpretability, and functional use. It also describes steering interventions that modify SAE-space representations to reduce Whisper hallucinations.

  • 3.1 SAE: SAEs represent each activation as a sparse linear combination of learned directions, using nonlinear encoders and reconstruction objectives with sparsity-related losses.The explored nonlinearities include Jump-ReLU, Top-k, and Batch-Top-k; Batch-Top-k was selected for subsequent experiments based on reconstruction quality and sparsity.
  • 3.2 Evaluation: The evaluation jointly measures reconstruction quality, feature stability, interpretability, and disentanglement rather than relying on a single metric.Reconstruction is assessed through L2–L0 trade-offs, while stability is tested across random seeds, layers, and model architectures.
  • 3.3 Feature robustness: Feature stability is measured by Intersection-over-Union of binary activation patterns, with coverage used to quantify transferable features across SAE sets.The comparisons span random seeds, layers, and HuBERT and Whisper architectures; within-SAE duplicates are identified through high IoU with another feature.
  • 3.3 Interpretability: Interpretability analysis combines domain specialization, classification-based feature ranking, manual inspection, semantic tests, label-based search, mel-spectrogram averaging, and caption-based analysis.The analyses operate at both frame and audio levels, including max-pooled feature aggregation and tasks such as gender, noise condition, accent, emotion, vowel, and phoneme analysis.
  • 3.4 Hallucination reduction: Steering applies a directional intervention in SAE space to bias Whisper activations away from hallucination-prone regions during inference.Features are selected using logistic-regression coefficients associated with hallucinations, and the intervention is controlled by a steering intensity parameter α.

4 Experiments

Experiments evaluate SAE reconstruction, stability, specialization, interpretability, feature semantics, neural alignment, and steering utility in Whisper and HuBERT. The results show stable, structured features that support selective analysis and reduce Whisper hallucinations.

  • Experimental setup: SAEs trained on every encoder layer of HuBERT and Whisper used diverse audio with online noise and music augmentation, BatchTopK, and L2 reconstruction loss.The downstream analysis focuses on HuBERT-base2 and Whisper-small.
  • SAE quality evaluation: Over 50% feature coverage across random seeds and below 5% within-checkpoint duplication indicate stable, low-redundancy representations.HuBERT and Whisper features do not align across models, while inter-layer coverage is high mainly in later layers.
  • Domain specialization: Whisper music features comprise roughly 20–28% of audio-level activations and peak around layer 5, while speech audio-level features peak near 13% before declining after layer 6.Frame-level speech specialization instead rises from about 2% at layer 6 to approximately 3.5% at layer 7.
  • Classification-based analysis: A small number of SAE features captures most classification information, with top-k curves saturating around k ≈10–150 for binary tasks and 500–3000 for complex multi-class tasks.Removing the information completely requires suppressing approximately 2000 features, consistent with redundancy and distributed encoding.
  • Feature interpretation: Final-layer phoneme classification reaches 0.92 for Whisper and 0.89 for HuBERT, while SAE features also identify paralinguistic and acoustic events with varying reliability.HuBERT layer 11 features 3249 and 3081 specialize in speech-boundary detection; auto-interpretation finds sounds such as alarms and birds but misses some phonetic details.
  • Hallucination reduction: Steering the top 100 SAE features reduces average non-speech false positive rate by 70%, from 0.37 to 0.11, with WER increasing only from 5.1% to 5.5%.Moderate steering strength provides the reported balance between false-positive reduction and recognition performance.
  • Neural alignment: Some SAE features significantly correlate with EEG activity at electrode Pz across time lags from 0 to 500 ms, with both positive and negative correlations.The authors note that near-zero-lag correlations can reflect access to left and right audio context.

5 Conclusion

The paper applies SAEs to HuBERT and Whisper to interpret their representations through robust, meaningful latent features. These features span broad domains and fine-grained audio concepts, support hallucination reduction, and correlate with EEG signals.

  • Conclusion: The work comprehensively applies SAEs to HuBERT and Whisper and introduces a cross-layer and cross-model evaluation metric with latent-analysis methods.The released analysis covers robustness, conceptual feature content, steering, and EEG correlation.
  • Conclusion: SAE features encode broad categories such as speech and music alongside phonemes, laughter, sighs, sneezing, and acoustically discovered properties.The conclusion characterizes these features as robust and meaningful.
  • Conclusion: Steering Whisper SAE features reduces hallucination false positives by 70%, while specific SAE features correlate with EEG signals.These findings provide the paper’s reported practical and neuroscientific applications.

Limitations

The evaluation is bounded by limited task and model coverage, captioning limitations for speech, and a narrow EEG analysis setup.

  • Downstream evaluation covers a limited set of classification tasks and applications, leaving broader audio-processing tasks for future work.Suggested extensions include speaker verification, speech enhancement, and audio generation.
  • The detailed analysis focuses on base and small model variants, while larger architectures and additional models were not comprehensively studied due to computational constraints.
  • The auto-interpretation method tends to produce generic captions for speech-related features and loses fine-grained phoneme-level information.This limitation is inherited from an audio captioning model trained primarily on music and sound data.
  • EEG correlation analysis is limited to a single Pz electrode and linear temporal response models.More comprehensive brain imaging and nonlinear modeling could reveal additional relationships.

A Extended SAE training details

The appendix details the model, dataset, architecture, and hyperparameter choices used to train the SAEs. A Batch-Top-K design with k = 50 and 8x expansion provides the reported balance among reconstruction, sparsity, and efficiency.

  • Base model Selection: The study trains SAEs on HuBERT-base, HuBERT-large, Whisper-small, and Whisper-largev3-turbo, but focuses downstream analysis on HuBERT-base and Whisper-small.EnCodec was excluded because its final-layer SAE did not learn a sufficiently sparse representation.
  • Dataset: The training corpus combines publicly available datasets with sampling weights designed to emphasize speech, music, and environmental sounds.High weights for MUSAN, FSD50K, and Nonspeech7k bias training toward non-speech audio, music, and environmental sounds.
  • SAE architecture Selection: Batch-Top-K was selected over Jump-ReLU and Top-K after preliminary analysis indicated slightly better reconstruction quality and sparsity control.All SAEs use an L2 reconstruction objective.
  • SAE-specific hyperparameters: The sweep varies expansion factors of 8x and 32x and sparsity levels k ∈25, 50, 75, 100, 200 across layers, with unit-norm input normalization.
  • Results: k = 50 with 8x expansion provides the best compromise among reconstruction fidelity, sparsity, and compression efficiency.The total number of active neurons still grows with expansion and remains at least twice the base model’s size, while smaller size–quality ratios are suboptimal.

B Feature robustness

Feature robustness is assessed through activation-based similarity, domain specialization, and coverage analyses across datasets, layers, and model conditions.

  • Feature characterization: Feature similarity is evaluated using activation frequency and average non-zero activation value at both frame and audio levels across speech, sounds, and music.
  • Domain assignment: Features are assigned to domains when one domain’s activation frequency exceeds the others by at least a progressive threshold; inactive features are labeled dead.Features failing every threshold are marked unassigned, and final labels aggregate three-way and pairwise domain combinations.
  • Datasets: The analysis uses seven datasets grouped into speech, sounds, and music categories to evaluate feature specialization.
  • Threshold analysis: Frame-level thresholds identify fine-grained specialization, whereas audio-level thresholds capture coarser patterns observable across complete samples.Audio-level thresholds of 0.5 and 0.3 emphasize substantial full-sample activation while preserving domain discrimination.

C.2 Frequency analysis

Frequency analysis compares activation frequency with average non-zero activation magnitude across Whisper and HuBERT layers 6–7. The models show distinct domain-specific activation profiles and specialization structures.

  • Audio-level frequency: Audio-level frequency spans 0 to 1.0 because sparse frame activations can occur across many samples.This aggregation produces broader sample-level coverage than frame-level frequency.
  • Whisper profile: Whisper music features cluster at activation magnitudes of approximately 3–6, while speech and sounds features are lower at approximately 0.5–2.The reported pattern suggests a dedicated set of high-magnitude music-responsive features in Whisper.
  • HuBERT profile: HuBERT shows dense specialization across frequencies up to 1.0 at comparable low activation magnitudes of approximately 0.4–1.0.Domain colors distinguish speech, sounds, and music without the activation-magnitude separation observed in Whisper.
  • Frame-level frequency: Whisper and HuBERT frame-level feature distributions concentrate most features below f_frame < 0.1, with substantial specialization near zero frequency.
  • Feature overlap: Across most layers, speech features are separated from sound and music features, while the sound set is nearly absorbed by the music set.
  • Model comparison: HuBERT has more audio-level speech features and fewer frame-level speech features than Whisper, consistent with different pre-training objectives and data composition.
  • Model comparison: Sound features are underrepresented in both models, nearly absent in HuBERT and sparse in Whisper, primarily at the frame level.The passage attributes this pattern to sound features co-occurring more often with music, subsuming them into music-associated activations.

E.2 Unlearning plots for various letters and regularization setups

The vowel unlearning experiments compare regularized and unregularized logistic regression to test how selectively phonetic information can be removed. Regularization makes apparent unlearning possible with far fewer removed features, while unregularized classifiers suggest more extensive removal is needed.

  • Experimental setup: The experiments use HuBERT’s 12th layer and compare standard LogisticRegression settings with unregularized settings across vowel classes.The regularized plots use default L2 penalty and C=1; unregularized plots use penalty='none'.
  • Regularization comparison: 160–400 features, or 3–6% of the total, suffice for comparable vowel unlearning with standard L2 regularization.Unregularized logistic regression requires removal of over 1000 features for successful unlearning.
  • Interpretive caveat: L2-regularized results may be overly optimistic because regularization can mask recoverable information by constraining classifier capacity.The authors therefore treat unregularized experiments as a more realistic assessment of true information removal.
  • Regularization comparison: Over 1000 features are required for successful unlearning with unregularized logistic regression.These classifiers provide a more extensive test of information removal but are difficult to use when features outnumber training samples.
  • Top-k probing: Adding features in Fisher-score order shows that one or two top-ranked features can restore high single-vowel classification accuracy.This indicates that the highest-ranked features carry enough phonetic information for reliable classification.

G Auto-interpretation details

Automatic interpretation identifies acoustic events and other features not represented in dataset annotations, but speech and phonetic features are often mischaracterized. The method’s outputs are limited by speech-heavy test data, threshold selection, and the caption model’s capabilities.

  • Label frequencies: Speech dominates the characteristic-label word map for both HuBERT and Whisper.The speech-heavy test set can make small but frequent activations favor speech interpretations over rarer music or sound events.
  • Interpretation limitations: Phonetic features may be mislabeled as generic multi-speaker dialogue because isolated phoneme clips lack broader acoustic context.A feature for the vowel sound “A” can therefore be confused by the captioning model.
  • Method limitations: The label-frequency analysis uses an empirically selected activation threshold of 0.1.This threshold is one of the method’s stated limitations.

H.2 Results and visualization

SAE-based steering selects informative hallucination-related features and balances false-positive reduction against speech-recognition quality. Moderate steering substantially lowers false speech detections, whereas aggressive steering harms the original ASR task.

  • ASR preservation: Steering vectors from Musan, FSD50k, and WHAM leave LibriSpeech WER essentially unchanged at around 0.05 across tested α values.This supports preservation of recognition quality on clean speech for these configurations.
  • Steering results: Top-100 SAE features with α=1 provide the preferred balance between hallucination suppression and ASR accuracy.The extreme top-100, α=3 setting achieves stronger suppression but significantly degrades the original task.
  • Cross-dataset behavior: A steering vector formed on FSD50k with k=100 and α=3 drives FPR close to zero across evaluation datasets.The authors distinguish this extreme setting from the lower-strength optimal setting used to preserve ASR accuracy.
  • Steering procedure: The SAE steering procedure jointly chooses α and k, where k is the number of SAE features selected by the hallucination classifier.Features are ranked by their importance in the classifier.
  • Feature interpretation: Speech-specific features are detected across multiple HuBERT and Whisper layers, including early and late network positions.The reported HuBERT layer-11 example is one instance of this broader distribution.

J Details of EEG experiments

The EEG analysis tests whether SAE features extracted from HuBERT and Whisper align with human neural responses during audiobook listening. A small fraction of features shows significant correlation with the Pz electrode response, sometimes at nearly zero lag.

  • EEG data: The study uses Pz EEG recordings from 19 subjects listening to five three-minute audiobook excerpts.Each participant contributes 15 minutes of listening data.
  • Preprocessing: SAE features are extracted from the final HuBERT-base and Whisper-base layers, while EEG signals are filtered to 1–8 Hz and both modalities are resampled to 128 Hz.Temporal-response functions are built with mTRFpy after normalization.
  • Feature selection: The analysis samples 1000 HuBERT and 1000 Whisper features that activate at least once per second on average.Feature-specific time lags are estimated on development data and tested on held-out data.
  • Correlation results: Around 1% of Whisper features and 1.5% of HuBERT features significantly correlate with Pz electrode responses at certain time lags.The analysis applies one-sided t-tests with Holm-Bonferroni correction at significance level 0.05.
  • Temporal interpretation: Nearly zero-lag correlations can occur because HuBERT and Whisper feature extractors access both left and right audio context.A feature may activate near the end of a word or sound even though the brain processes the audio over time.
Loading 2602.05027v2…