Fairer speech recognition improves accuracy for overlapping demographic groups
InterCorrect: Intersection-Aware Correction of Demographic Model Merging for Fair ASR
Computation and LanguageSound
Summary
Speech recognition systems often make more mistakes for certain groups of people, especially for those who belong to multiple groups at once. The authors studied how to improve these systems by fine-tuning parts of the model for specific groups and then combining them carefully. They also added special corrections for people at the intersection of these groups to reduce errors further. Their method lowered the overall word error rate and improved fairness across different demographics, though smaller error rates don’t always mean all groups are treated equally.
What this means in practice
- •For speech recognition developers: Build ASR systems that better handle speakers from multiple demographic groups by merging subgroup-specific models with intersection corrections.
- •For voice assistant teams: Improve voice assistant accuracy and fairness by incorporating intersection-aware corrections into demographic-adapted speech recognition models.
Authors
Ashley E. Bravo-Bravo, Yuchen Zhang, Haralambos Mouratidis, Ravi Shekhar, Monorama Swain
Abstract
Automatic Speech Recognition (ASR) systems often show uneven performance across demographic groups, and errors can be especially difficult to address for speakers belonging to multiple demographic groups. This work studies demographic-aware model merging for fair Speech-LLM-based ASR. Starting from a SLAM-ASR-based model, we fine-tune only the connector on demographic-specific subsets and merge the resulting subgroup-adapted connectors into a global model. We then identify critical cross-axis demographic pairs using subgroup WER and task-vector conflict, and apply intersection-specific correction vectors to the global merged model. Experiments on Fair-Speech show that global demographic merging improves overall WER over the base model, while intersection correction provides additional gains for several merging strategies. In particular, TIES with WER-based correction achieves the best overall WER, reducing it from 7.38\% to 5.13\%. Subgroup and disparity analyses further show that the proposed approach improves performance across demographic axes, while highlighting that lower average WER does not always imply reduced subgroup disparity.