Voice activity detection in noisy environment using dynamic changes of speech in a modulation frequency range of 1 to 16 Hz
Suci Dwijayanti, Masato Miyoshi · The Journal of the Acoustical Society of America · 2016
Voice activity detection (VAD) is usually placed at the preprocessing stage of various speech applications. There have been a lot of studies to find acoustic features that are effective in distinguishing speech/non-speech segments on a recorded signal. In this report, we utilize speech characteristics in a modulation frequency range of 1-to-16 Hz, which seem beneficial to avoid misjudgment. The signal amplitudes are evaluated in common logarithms not to miss small changes in speech energy that may have relations with the starting or ending point of an utterance. In order to capture such changes, the first and second derivatives of the signal are calculated. In finding the starting point, the positive second derivatives seem effective. In finding the ending point, the combination of the negative first derivatives and the positive second derivatives seem effective. And these features should be calculated for individual subbands. On these conditions, Deep Neural Network (DNN) is successfully trained to determine the speech/non-speech segments. In this report, we will evaluate this method in comparison with conventional ones for various noisy utterances.