Electromyography (EMG) offers a non-invasive way to turn silently articulated speech into text, but the approach has been held back by a lack of paired EMG-text training data. Without enough examples, models struggle to map muscle signals reliably to words.
PhonoBridge, introduced in a new arXiv preprint, is designed to tackle this limitation. The name suggests a phoneme-mediated approach: it likely breaks speech into intermediate phonetic units before reconstructing text, and the title's "error-aware" component hints at a mechanism for handling mistakes in that process.
The work is still at the preprint stage, and the abstract does not provide quantitative results. But the core idea—reducing dependence on large paired datasets—could make silent speech interfaces more practical for assistive technology and other applications where EMG data is scarce.