Robust voice activity detection based on LSTM recurrent neural networks and modulation spectrum
Phuttapong Sertsi, Surasak Boonkla, Vataya Chunwijitra, Nattapong Kurpukdee, Chai Wutiwiwatchai · 2017
Voice activity detection (VAD) used for classifying speech/non-speech sections of a speech signal still suffers from noisy environments. In this paper, we cooperate the modulation spectrum (MS) and the long short-term memory recurrent neural network (LSTM) to improve the robustness. The baseline LSTM used conventional speech features in training and classifying speech and non-speech sections. The proposed VAD system by using MS as another speech feature in order to increase the robustness. In addition, we propose a new approach for the computation of the MS feature. The results showed that the accuracy of using the proposed method can be improved under both seen and unseen noise conditions compared with the baseline technique.