Near-Infrared Neck Sensing for Silent Speech Classification: A Feasibility Study Combining Video and Temporal Signal Analyses
Abstract
Silent speech interfaces offer a route to private and assistive communication without depending on an audible signal. This feasibility study investigates whether near-infrared sensing on the anterior neck contains class-discriminative information that could support future silent speech decoding. We combine a wide-field video stage with a localized reflective sensing stage. Video recordings were analyzed using region-wise image trajectories, dynamic time warping, and a CNN-based classifier, while demodulated optical signals were assessed using temporal DTW analysis and a 1D CNN classification task. Across both modalities, the lower central neck region showed strong discriminability. The best video-based classifier achieved 84.75% accuracy, while the signal-based classifier achieved 56% balanced accuracy under the current task definition. Overall, the results support near-infrared neck sensing as a feasible route for silent-speech interfaces, while indicating that future systems may benefit from fused designs combining wide-field video information with high-bandwidth local optical sensing.