Emerging Interfaces
PhonoBridge Reads Silently Mouthed Words Off the Face, at a 24.77% Error Rate
Silent-speech interfaces are starved of data: every hour of training requires someone wearing electrodes and mouthing words. PhonoBridge routes EMG through phonemes first, which lets a language model learn from ordinary speech transcripts instead — and cuts word error rate from 28.27% to 24.77%.