Representation Steering Fails to Reduce ASR Word Error Rate Gaps
September 17, 2026
Injecting probed speaker attributes into Whisper, HuBERT, and Wav2Vec2 encoders does not effectively reduce word error rate (WER) disparities across demographic groups. While attributes like sex and accent are highly decodable, steering representations fails to yield consistent downstream fairness improvements.
HOW THIS AFFECTS YOU
●
researcherInternal representation steering may be an insufficient mechanism for addressing systemic bias in speech models.
●
policyBe cautious of technical fixes that claim to solve demographic bias via representation manipulation.