VibraGlot: Contactless Vocal Disorder Screening via mmWave Neck-Surface Vibration Sensing

Vocal disorders often begin as subtle mechanical irregularities in vocal-fold vibration before speech quality noticeably degrades. Early identification of such abnormalities is valuable because timely intervention may prevent progression to more advanced stages. However, detecting these weak early cues outside clinical settings remains challenging: microphone-based analysis observes an acoustically shaped signal that may only weakly express source-level irregularity, while contact-based vibration sensors suffer from low user compliance. We present VibraGlot, a contactless screening system that leverages millimeter-wave radar to sense neck-surface micro-vibrations during a short, non-linguistic sustained phonation (3-5 s). Rather than adopting generic end-to-end learning, VibraGlot uses a kinematics-first design. It first isolates the laryngeal region through observability-aware locking to suppress static clutter, then performs cycle-synchronous reconstruction in a canonical phase domain. Crucially, this pipeline explicitly preserves template-violating motion as structured residuals, transforming irregular mechanical behavior into diagnostic features rather than smoothing it away as noise. We evaluate VibraGlot on a clinical cohort of 74 subjects spanning five common vocal pathologies and healthy controls, collected in both hospital and home environments. Under a strict minimal-input protocol, the system achieves a macro-F1 score exceeding 0.91. These results demonstrate that radar-based neck-surface kinematics provide a robust, interpretable, and contactless pathway for scalable early vocal disorder screening, particularly when pathology is still expressed as subtle mechanical irregularity.