Beyond Transcription: Mechanistic Interpretability in ASR

TL;DR

This paper adapts interpretability methods to analyze how acoustic and semantic information evolves in ASR systems, revealing internal dynamics and biases that can inform improvements in transparency and robustness.

Contribution

It systematically applies established interpretability techniques to ASR, uncovering new insights into internal representations and interactions within speech recognition models.

Findings

Identified encoder-decoder interactions causing repetition hallucinations

Discovered semantic biases in deep acoustic representations

Revealed internal dynamics of information flow in ASR models

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.