Improved Voice Activity Detector Based on Stacking Weak Classifiers

Abstract

The article considers the problem of speech fragments extraction. The authors have brought together a dataset of speech signals VADSpeakersDB. This dataset consists of phonograms recorded with the help of an application for video conferences. The research considers an original algorithm based on the stacking of independent speech activity detectors and compares it with traditional approaches. The solution has a high accuracy of detecting voice fragments-more than 90%.