PhonoBridge preprint reports lower error rate in silent-speech decoding

Researchers report that PhonoBridge, a system for converting silently articulated speech into text using muscle signals, achieved a 24.77% word error rate on the Open Vocabulary silent-speech dataset. The preprint compares that result with the previously reported best of 28.27% on the same test split. The system first converts electromyography signals into phonemes, then uses a large language model to reconstruct text. It expands training data using phoneme-text pairs derived from LibriSpeech transcripts without additional muscle recordings, introducing errors that mimic the frontend's mistakes. The reported improvement is limited to the dataset's test split.