Temporal Taxation Compounds Under Post-Training Compression of Whisper Models
Source
Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy
Author
Srishti Ginjala, Eric Fosler-Lussier, Christopher W. Myers, Srinivasan Parthasarathy
Date
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 quantizationShrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.Full definition → multiplied catastrophic transcript loops on West African accents by five to seven times. distillationTraining a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.Full definition → 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.
Why it matters
If you're compressing a speech model for edge deployment, fairness audits done at full precision don't transfer; compression itself can double the transcription-quality gap between demographic groups.
articleRepresented but Ignored: A Causal Account of Prosodic Underuse in Audio-Language ModelsLinkai Peng, Baorian Nuchged
videoSpeech-to-Speech Model Research at Google DeepMind — Valeria Wu Fon & Tom Ouyang, Google DeepMindAI Engineer
articleVoice Memory for Agentic Speech RecognitionChao-Han Huck Yang, Zih-Ching Chen, Piotr Zelasko, Zhehuai Chen, Jagadeesh Balam, Boris Ginsburg