A Method Based on Support Vector Machine for Voice Activity Detection on Isolated Words

Abstract

Voice activity detection (VAD) is an essential part of speech processing system, which aims at extracting active speech information from continuous speech signals. In this paper, we firstly treat VAD as a two-class problem: active speech frame and nonspeech frame, in which the active speech frame is composed of information frame and noise frame. Support vector machine is then applied to solve the two-class problem. Finally, noise frames are removed from the active speech frames by the prior knowledge of noise. An experiment on a speech dataset of isolated numbers shows the accuracy and recall rate on test set reaches 99.49% and 98.06% respectively, which indicates that the proposed method is effective for isolated-word VAD task.