
One speech-to-speech model got a speaker’s gender wrong 2% of the time when voice and topic matched a stereotype. When they clashed, it got it wrong 90% of the time. We tested five models on clips that paired a male or female voice with a stereotypically male, female, or neutral topic. Their output voices stayed consistent. How they referred to the speaker didn’t. One model called a woman “he” after she talked about rebuilding an engine. If you build or buy these models, test cases where voice and topic disagree. #NeurIPS2026 workshop “Can We Trust AI Evaluation?”, Sydney, Dec 12. arxiv.org/abs/2609.09263 Co-authored by @harshitr (Centific) and Tanu Mitra (University of Washington)


















