Two-step Judgment Algorithm for Robust Voice Activity Detection Based on Deep Neural Networks

Voice Activity Detection (VAD) is an important front-end process for speech-based applications such as automatic speech recognition (ASR) and speaker diarization. VAD attempts to identify all the segments containing speech in an audio signal. In this paper, a robust VAD system is developed based on deep neural network (DNN) fusion with Combo-SAD. DNN model is an effective supervised approach that can achieve 4% of missed detection rate (Pmiss) at a falsealarm rate (Pfa) of 5%, Combo-SAD is an unsupervised approach designed for noise robust and reported a 5% Pmiss at Pfa of 3%. Combining the advantages of both techniques, this paper attempts to design a 2-step judgment approach. Experimental results on database containing various type of audios show that the overall error rate reaches 13.50%, which indicates the proposed VAD system is robust and effective.