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

Researcher Swapnil Parekh has introduced ACES (Accent Subspaces for Coupling, Explanations, and Stress-Testing), a novel diagnostic framework for auditing bias in Automatic Speech Recognition systems. The research, published on arXiv, tackles the persistent problem of performance disparities across different English accents in models like Meta's Wav2Vec2-base. ACES works by extracting low-dimensional subspaces within the model's internal representations that are most sensitive to accent information, providing a precise tool to probe where and how bias manifests, moving beyond simple output metrics to understand internal mechanisms.

The technical analysis of Wav2Vec2-base across five English accents revealed that accent information is surprisingly concentrated in a very small part of the network—specifically an 8-dimensional subspace in the model's third layer. The magnitude of a speech sample's projection onto this 'accent subspace' showed a measurable correlation (r=0.26) with its word error rate (WER). Crucially, the study found that attempts to linearly attenuate or 'erase' this subspace did not reduce performance disparity and even slightly worsened it, indicating that accent-relevant features are fundamentally entangled with the cues needed for accurate speech recognition. This positions ACES as a vital diagnostic tool for developers, highlighting that fairness interventions must be more sophisticated than simple feature removal.