mlsanityData to 14 Oct 2026

Beyond Transcription: Mechanistic Interpretability in ASR

arXiv 2508.15882 ↗First posted 21 Aug 2025cs.SD cs.CL cs.LGcited by 13

On the map: Mechanistic interpretability → Knowledge and representations inside LLMs

Abstract (arXiv)

Interpretability methods have recently gained significant attention, particularly in the context of large language models, enabling insights into linguistic representations, error detection, and model behaviors such as hallucinations and repetitions. However, these techniques remain underexplored in automatic speech recognition (ASR), despite their potential to advance both the performance and interpretability of ASR systems. In this work, we adapt and systematically apply established interpretability methods such as logit lens, linear probing, and activation patching, to examine how acoustic and semantic information evolves across layers in ASR systems. Our experiments reveal previously unknown internal dynamics, including specific encoder-decoder interactions responsible for repetition hallucinations and semantic biases encoded deep within acoustic representations. These insights demonstrate the benefits of extending and applying interpretability techniques to speech recognition, opening promising directions for future research on improving model transparency and robustness.

Key facts. Cited by 13 papers in the map (data through 14 Oct 2026). Placed in Knowledge and representations inside LLMs, one of 13 directions in Mechanistic interpretability. Builds on 32 of its 51 references that are in the map. First posted on arXiv 21 Aug 2025.

Builds on

32 of 51 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Do Multilingual LLMs Think In English?21 Feb
  2. 2023Eliciting Latent Predictions from Transformers with the Tuned Lens14 Mar

Other references · 30

  1. 2017Attention Is All You Needcited by 60,361
  2. 2020Language Models are Few-Shot Learnerscited by 32,419
  3. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 14,132
  4. 2022Robust Speech Recognition via Large-Scale Weak Supervisioncited by 4,057
  5. 2020wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representationscited by 3,905
  6. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  7. 2016Enriching Word Vectors with Subword Informationcited by 1,583
  8. 2019Common Voice: A Massively-Multilingual Speech Corpuscited by 1,384
  9. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  10. 2023Qwen-Audio: Advancing Universal Audio Understanding via Unified Large-Scale Audio-Language Modelscited by 608
  11. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 532
  12. 2020MLS: A Large-Scale Multilingual Dataset for Speech Researchcited by 476
18 more
  1. 2015MUSAN: A Music, Speech, and Noise Corpuscited by 425
  2. 2017AISHELL-1: An Open-Source Mandarin Speech Corpus and A Speech Recognition Baselinecited by 373
  3. 2020FSD50K: An Open Dataset of Human-Labeled Sound Eventscited by 342
  4. 2024LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinationscited by 244
  5. 2024How to use and interpret activation patchingcited by 214
  6. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 209
  7. 2023Linearity of Relation Decoding in Transformer Language Modelscited by 200
  8. 2023Localizing Model Behavior with Path Patchingcited by 195
  9. 2024AtP*: An efficient and scalable method for localizing LLM behaviour to componentscited by 111
  10. 2024Transformers need glasses! Information over-squashing in language taskscited by 91
  11. 2023Whisper-AT: Noise-Robust Automatic Speech Recognizers are Also Strong General Audio Event Taggerscited by 72
  12. 2025Detecting High-Stakes Interactions with Activation Probescited by 53
  13. 2024Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelinescited by 30
  14. 2024Hallucinations in Neural Automatic Speech Recognition: Identifying Errors and Hallucinatory Modelscited by 27
  15. 2024Do Prompts Really Prompt? Exploring the Prompt Understanding Capability of Whispercited by 12
  16. 2025A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinationscited by 9
  17. 2024Whisper-PMFA: Partial Multi-Scale Feature Aggregation for Speaker Verification using Whisper Modelscited by 3
  18. 2023Speaker and Language Change Detection using Wav2vec2 and Whispercited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 13.

13 papers.

Sort:

By area:

  1. 2026Geometric Alignment without Functional Equivalence: A Layer-wise Analysis of the Speech-Text Modality Gapcited by 7
  2. 2026Are Audio-Language Models Listening? Audio-Specialist Heads for Adaptive Audio Steeringcited by 4
  3. 2026Mechanistic Interpretability of Brain-to-Speech Models Across Speech Modescited by 1
  4. 2026From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detectioncited by 1
  5. 2026Scalable Music Cover Retrieval Using Lyrics-Aligned Audio Embeddingscited by 0
  6. 2026Demonstration of Adapt4Me: An Uncertainty-Aware Authoring Environment for Personalizing Automatic Speech Recognition to Non-normative Speechcited by 0
  7. 2026HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systemscited by 0
  8. 2026LOPA: Enhancing Spoken Language Assessment via Latent Ordinal Prototype Alignmentcited by 0
  9. 2026CASA: Content-Acoustic Speaking Assessment with Speech Encoder and Large Language Modelcited by 0
  10. 2026Towards Quantifying Benchmark Optimization in ASR Modelscited by 0
  11. 2026The Anatomy of an ASR Hallucinationcited by 0
  12. 2026Mechanistic Interpretability Reveals Shared Causal Subspaces in Brain-to-Speech Decoderscited by 0
1 more
  1. 2026Tracing a Sparse Emotion-Control Circuit in LLM-Based Text-to-Speechcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025TRACE: Training and Inference-Time Interpretability Analysis for Language Models
  2. 2024Towards Unifying Interpretability and Control: Evaluation via Intervention
  3. 2025Behind the Scenes: Mechanistic Interpretability of LoRA-adapted Whisper for Speech Emotion Recognition
  4. 2026Scaling Inherently Interpretable Language Models
  5. 2025Explainable Disentanglement on Discrete Speech Representations for Noise-Robust ASR
  6. 2026MechaTerp-TRACE: A Novel Approach for Component Ablation Analysis in Language Models
  7. 2026On the Interpretability of Whisper Encodings Using Sparse Autoencoders
  8. 2026Mechanistic Interpretability for Large Language Model Alignment: Progress, Challenges, and Future Directions