SilentVoice: Unnoticeable Voice Input by Ingressive Speech
summary
A wearable microphone sits almost flush to the mouth and uses ingressive speech (inhaling while shaping the mouth) to create a whisper-like sound source, achieving ultra-low leakage (under 39dB(A) leakage) and allowing very close mic placement (<1 mm). Airflow sensing separates commands with 98.8% activation accuracy. Off-the-shelf recognizers start at 83% WER on this input but adapt to 1.8% WER speaker-dependent (7.0% speaker-independent) and sustain ~6.9% WER on 80 dB subway noise.
AI-written summary · about