Restoring natural conversational speech in anarthric patients paralyzed by brainstem strokes requires decoding intended vocal tract kinematics directly from speech sensorimotor cortex. This study evaluates deep autoregressive vocoders running on low-latency edge accelerators with conversational latency bounds under 85 milliseconds.
Electrocorticography (ECoG) grids placed over ventral sensorimotor cortex stream high-gamma (70-150 Hz) power to predict pitch, formants, and tongue articulation positions.
Streaming convolutional vocoders synthesize natural audio speech waveforms chunk-by-chunk, allowing patients to maintain natural conversational pacing without awkward computational pauses.