Beyond Transcription: Mechanistic Interpretability in ASR

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Glazer, Neta, Segal-Feldman, Yael, Segev, Hilit, Shamsian, Aviv, Buchnick, Asaf, Hetz, Gill, Fetaya, Ethan, Keshet, Joseph, Navon, Aviv
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866915456607059968
author Glazer, Neta
Segal-Feldman, Yael
Segev, Hilit
Shamsian, Aviv
Buchnick, Asaf
Hetz, Gill
Fetaya, Ethan
Keshet, Joseph
Navon, Aviv
author_facet Glazer, Neta
Segal-Feldman, Yael
Segev, Hilit
Shamsian, Aviv
Buchnick, Asaf
Hetz, Gill
Fetaya, Ethan
Keshet, Joseph
Navon, Aviv
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2508_15882
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Beyond Transcription: Mechanistic Interpretability in ASR
Glazer, Neta
Segal-Feldman, Yael
Segev, Hilit
Shamsian, Aviv
Buchnick, Asaf
Hetz, Gill
Fetaya, Ethan
Keshet, Joseph
Navon, Aviv
Sound
Computation and Language
Machine Learning
Audio and Speech Processing
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.
title Beyond Transcription: Mechanistic Interpretability in ASR
topic Sound
Computation and Language
Machine Learning
Audio and Speech Processing
url https://arxiv.org/abs/2508.15882