Compression of speech models increases fairness gaps for some groups
Temporal Taxation Compounds Under Post-Training Compression of Whisper Models
Computation and LanguageSound
Summary
Speech recognition models are often made smaller or faster before being used, but these changes can make them less fair for certain groups of people. The authors found that one type of model shrinking, called pruning, makes errors happen more often for some racial groups compared to others, meaning those people would need to spend more time fixing mistakes. Another shrinking method, called distillation, tends to reduce these fairness gaps. The study shows it’s important to check fairness after making these smaller models, not just before.
What this means in practice
- •For speech technology engineers: Adjust post-training compression processes to better manage fairness across demographic groups in deployed speech recognition systems.
- •For mobile device developers: Optimize model compression methods to reduce error burdens for users with diverse accents when running speech recognition locally.
Authors
Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy
Abstract
Automatic speech recognition models are audited for demographic fairness at full precision, yet the models that ship to production have been quantized, pruned, and distilled. We ask whether post-training weight compression, which alters model weights rather than the audio signal or its feature representation, redistributes error burden across demographic groups. Across the Whisper family on Fair-Speech, Common Voice 25, and AfriSpeech-200, 50% Wanda pruning of Whisper-large-v3 sharply widens the Black/AA-vs-Asian temporal-taxation differential on Fair-Speech: the absolute word-error-rate gap between the worst- and best-served groups more than doubles; at an assumed cost of five seconds of correction effort per transcription error this is a rise from 30 to 64 seconds of correction time per minute of speech. This +111% relative increase is invariant to the assumed per-error cost, survives an audio-quality control, and is only partly mitigated by beam-search decoding, which still leaves an +86% increase. At edge model size, INT4 HQQ quantization compounds catastrophic transcript loops on West African accents by factors of five to seven. Distillation, by contrast, narrows demographic gaps in 21 of 27 evaluated settings (teacher-student pair, precision, and dataset), with the exceptions concentrated on a single model pair. We cast the temporal-taxation construct of Choi and Choi (2025) as a quantitative metric, and show that single-snapshot fairness audits on full-precision models do not capture the deployment-time burden that compression places on already-marginalized speakers.