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

Researchers introduce ACES, a new method to analyze how automatic speech recognition systems perform differently across accents. The study finds that accent information is concentrated in early neural network layers and is deeply intertwined with speech recognition capabilities, making simple bias removal ineffective.

• →ACES method identifies accent-discriminative subspaces in ASR models to understand performance disparities across different English accents.
• →Accent information concentrates in low-dimensional early-layer subspaces of Wav2Vec2 models (layer 3, k=8).
• →Projection magnitude in accent subspaces correlates with word error rates, indicating connection to model performance.
• →Subspace-constrained perturbations show stronger coupling between representation changes and performance degradation than random controls.
• →Simple linear attenuation of accent subspaces does not reduce bias and may worsen performance, suggesting deep entanglement with recognition features.