ACES: Accent Subspaces for Coupling, Explanations, and Stress-Testing in Automatic Speech Recognition

Swapnil Parekh · 2026

Read ACES: Accent Subspaces for Coupling, Explanations, and Stress-Testing in Automatic Speech Recognition by Swapnil Parekh (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03359.txt.

Description

Swapnil Parekh audits speech recognition for accent bias, asking whether accuracy gaps across accents run deep or sit on the surface. His 2026 audit, ACES, works in three stages on a model's internal representations. It first finds the directions in representation space that carry accent information, then aims adversarial attacks along those directions, and finally tests whether deleting them would help fairness.

On a standard wav2vec model covering seven accents, tiny inaudible perturbations along the accent directions widened the error-rate gap by nearly half, far beyond random-direction controls. A label-shuffling check confirmed the effect tracks real accent structure. Surprisingly, partly removing the accent directions made both accuracy and disparity worse. Accent information is deeply entangled with recognition itself, so Parekh positions the subspaces as auditing tools, not simple erasure levers.

Catalog record

CREATORSwapnil Parekh
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21106
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03359

Browse the full library · About the Tower of Babel