Intel / article

Temporal Taxation Compounds Under Post-Training Compression of Whisper Models

Source: arxiv.org
Author: Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy
Date: Published Sep 25, 2026

Context

Compressing a speech-recognition model can change which speakers bear its errors. In a Whisper study, removing weights through 50% Wanda pruning more than doubled the word-error gap between Black/African American and Asian speakers in the Fair-Speech dataset.

The researchers translated that gap into estimated correction work: about 30 versus 64 additional seconds per minute of speech, assuming five seconds to repair each error. These are calculated time costs, not timed editing sessions. The result shows why a fairness assessment of the original model may not describe the compressed version people actually use.

Key takeaways · AI-distilled

50% Wanda pruning of Whisper-large-v3 more than doubled the absolute word-error-rate gap between the worst- and best-served groups on Fair-Speech (Black/AA vs Asian speakers), a +111% relative increase.
At an assumed five seconds of correction per error, that gap grows from 30 to 64 seconds of correction time per minute of speech. The relative increase does not depend on the assumed cost and survives an audio-quality control.
Beam-search decoding only partly mitigates the damage: the increase still stands at +86%.
At edge model sizes, INT4 HQQ multiplied catastrophic transcript loops on West African accents by five to seven times. went the other way, narrowing demographic gaps in 21 of 27 settings, with exceptions concentrated on one model pair.

Terms in this piece · Glossary

quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.

Read the source arxiv.org