Beyond Transcription: Mechanistic Interpretability in ASR
Proceedings of the AAAI Conference on Artificial Intelligence · 14.03.2026 · 40 (44) · 37407-37416
DOI: 10.1609/aaai.v40i44.41073 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence, online erschienen am 14.03.2026. Verfasst von Neta Glazer und 8 weiteren Autor:innen.
Abstract (Originalsprache)
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.
Zitieren
APA
Glazer, N., Segal-Feldman, Y., Segev, H., Shamsian, A., Buchnick, A., Hetz, G., Fetaya, E., Keshet, J., & Navon, A. (2026). Beyond Transcription: Mechanistic Interpretability in ASR. Proceedings of the AAAI Conference on Artificial Intelligence, 40(44), 37407-37416. https://doi.org/10.1609/aaai.v40i44.41073
Vancouver
Glazer N, Segal-Feldman Y, Segev H, Shamsian A, Buchnick A, Hetz G, et al. Beyond Transcription: Mechanistic Interpretability in ASR. Proceedings of the AAAI Conference on Artificial Intelligence. 2026;40(44):37407-37416. doi:10.1609/aaai.v40i44.41073
BibTeX
@article{Glazer2026,
author = {Glazer, Neta and Segal-Feldman, Yael and Segev, Hilit and Shamsian, Aviv and Buchnick, Asaf and Hetz, Gill and Fetaya, Ethan and Keshet, Joseph and Navon, Aviv},
title = {Beyond Transcription: Mechanistic Interpretability in ASR},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
year = {2026},
volume = {40},
number = {44},
pages = {37407--37416},
doi = {10.1609/aaai.v40i44.41073}
}
EUVIMED