Beyond Transcription: Mechanistic Interpretability in ASR

Neta Glazer, Yael Segal-Feldman, Hilit Segev, Aviv Shamsian, Asaf Buchnick, Gill Hetz, Ethan Fetaya, Joseph Keshet, Aviv Navon

Abstract

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.

Introduction

While modern Automatic Speech Recognition (ASR) systems like Whisper and Qwen2-Audio have reached incredible levels of performance, they remain largely "black boxes." We know they work, but we do not fully understand how they represent information internally or why they occasionally fail with bizarre behaviors like "repetition loops" (repeating the same word indefinitely) or "hallucinations" (generating text from silence).

Most mechanistic interpretability research has focused on Large Language Models (LLMs). This research bridges the gap by applying reverse-engineering techniques to the audio-to-text pipeline, aiming to understand the internal dynamics of encoder-decoder ASR architectures and identify the specific components responsible for both successful transcription and systematic failure.

Methodology

To pull back the curtain on these models, the researchers adapted several established interpretability techniques to the ASR domain:

  • Logit Lens: Projecting the internal "residual stream" of the decoder onto the final vocabulary to observe how predictions evolve layer-by-layer.
  • Activation Probing: Training simple linear classifiers on frozen internal representations to determine if specific attributes—such as speaker gender, accent, or semantic category—are "linearly decodable."
  • Component Patching & Ablation: Swapping or zeroing out specific internal activations (like a single attention head) to observe the causal effect on the output.
  • Encoder Lens: A novel technique where intermediate encoder layers are passed directly to the decoder (skipping the top encoder layers) to visualize what the encoder "thinks" at different depths.

These methods were tested across a diverse array of datasets, including LibriSpeech, Common Voice, and specialized sets for noise (MUSAN) and semantic categorization.

The Encoder as a Semantic Engine

A primary finding of this study is that ASR encoders do far more than just process acoustic features. Contrary to the belief that the encoder is a simple "ear," the researchers found that Whisper’s encoder builds a complex semantic hierarchy. By the middle layers, it starts behaving like a "loosely grounded language model," forming conceptual clusters before the decoder even gets involved.

Linear probes reached an 85.6% average accuracy in classifying semantic categories (e.g., "Fruits" vs. "Countries") from encoder activations, with some pairs reaching 100% accuracy. Furthermore, probes successfully extracted speaker gender (94.6%), environment noise (90.0%), and accents (97.0%) from intermediate layers, proving that the encoder maintains a rich, multi-faceted representation of the audio signal.

Mechanistic Drivers of Failure

One of the most significant contributions of this work is the identification of specific "circuits" responsible for model failures:

Hallucinations

The research reveals that the model effectively "knows" when it is hallucinating. By probing the decoder’s residual stream at the final <eos> token, the researchers could predict if a transcription was a hallucination with 93.4% accuracy. The decoder residual stream can also perfectly distinguish (100% accuracy) between actual speech and non-speech audio (like white noise or music) being transcribed as words.

Repetition Loops

Repetition loops are not global model failures but are often driven by specific, localized components. In Whisper, the researchers identified that intervening on just one specific attention head (Layer 18, Head 13) could suppress repetitions by 78.1%. This suggests that targeted "model surgery" could fix these behaviors without the need for expensive retraining.

The Memorization Trap

Large Audio-Language Models like Qwen2-Audio sometimes ignore audio entirely in favor of memorized training data. For example, a specific phrase ("Kids are talking by the door") appeared in the output of 390 out of 400 files when probing early layers, regardless of the actual audio input. This highlights a significant challenge in scaling ASR models: the risk of semantic priors overriding acoustic evidence.

Conclusion

This study demonstrates that mechanistic interpretability is a powerful tool for ASR research. By revealing that encoders act as semantic engines and that hallucinations are internally detectable, this work opens the door for real-time monitors that flag errors during inference and fine-grained editing techniques to fix model biases. While the study focuses on Whisper and Qwen2-Audio, the methodologies provide a blueprint for making all speech-to-text systems more transparent and robust.

Summary updated

Abstract

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.
BibTeX
@inproceedings{glazer2026beyond,
  author = {Neta Glazer and Yael Segal-Feldman and Hilit Segev and Aviv Shamsian and Asaf Buchnick and Gill Hetz and Ethan Fetaya and Joseph Keshet and Aviv Navon},
  title = {{Beyond Transcription: Mechanistic Interpretability in ASR}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i44.41073},
}
Figure