Researchers build predictor for unreliable voice queries

Researchers report a way to detect when a speech-based AI model is likely to mishear a user before it generates an answer. Their preprint says asking the model to judge its own transcription works poorly; a small predictor reading signals from its audio encoder performs better without changing the underlying model. The predictor scored 81.10% in-domain and 78.09% cross-domain macro-F1, exceeding the strongest baselines by 10.33 and 11.93 points, respectively. It could prompt users to clarify doubtful queries while allowing reliable ones to proceed. The results are from the researchers' evaluation, not a deployed service.