Supervised Optimization Improves ASR Performance and Demographic Fairness
October 21, 2026
Supervised optimization of steering vectors effectively improves ASR performance and demographic fairness across gender, age, and dialect. While paired activation subtraction fails to improve ASR performance, vectors trained on acoustically similar speakers yield competitive or superior results compared to training on the same speaker.
HOW THIS AFFECTS YOU
●
builderYou can improve ASR fairness and accuracy without retraining the base model by applying optimized steering vectors.
●
researcherYou can use supervised optimization rather than subtraction to derive more effective steering vectors.