HiVAD : A Voice Activity Detection Application Based on Deep Learning

Abstract

In this paper, the detection of sound activity is presented on smartphones in realtime with convolutional neural networks. Reduced computing time is a problem from previous studies. Despite the use of machine learning approaches, there are still many shortcomings from previous research. A log-mel energy spectrogram narrates the sound signal image. Then the sound signal image is inputted into CNN's deep learning to classify the human voice and noise. HiVAD outperformed the percentage of other VAD methods, namely G729B, Sohn, and RF from the test results shown with an average SHR accuracy of 15.89%, 28.98%, 42.13% at 0dB, 8.67%, 16.29% ,17.63% at 5 dB, and 1.35%, 7.72%, 5.14% at 10 dB. In addition, the Multi-threading mechanism enables efficient computing for real-time. This study shows that CNN's architecture on HiVAD significantly improves the accuracy of sound activity detection.