Skin-Conformal Inertial Sensing Interface for Continuous Silent Speech Recognition

Abstract Silent speech interfaces offer a promising route for restoring communication without phonation; however, existing approaches are often limited by environmental sensitivity, rigid sensor mechanics, and dependence on fixed sentence classes. Here, we present a skin-conformal inertial sensing interface for continuous silent speech recognition from lip articulation. Commercial inertial measurement units are integrated onto a soft substrate through laser-patterned oxidized gallium–indium interconnects, enabling stable electrical performance under mechanical deformation and intimate coupling to dynamic facial skin. This conformal architecture mitigates motion artifacts to ensure high signal fidelity during dynamic articulation and supports robust operation in acoustically noisy and visually constrained environments. Combined with a data-efficient sliding-window decoding strategy, the system generalizes from isolated-word training to recognition of previously untrained sequences of different lengths composed of words from the trained vocabulary, with an average accuracy of 92.5%. Discrete word recognition reaches 99.6% across a 25-word daily vocabulary. These results establish a wearable silent speech sensing platform that supports continuous fixed-lexicon communication for assistive human–machine interaction.