Robust and Preceding Speech Detection Using EMG

EMG-based VAD is proposed. It is generally known that EMG signals are not affected by acoustic noise and preceding voice. Robust but preceding detection of speech can be expected by using these features. The proposed method extracts one VAD parameter to judge speech or silence from among n EMG signal channels. Experiments are performed to confirm the effectiveness of the proposed method. The results show that some EMG signals reliably precede voice by 30 ms or more. Compared with conventional audio VAD, the proposed method seldom drops the speech head and the method´s performance is relatively insensitive to background noise