Voice Activity Detection Using Weighted K-Means Thresholding Algorithm

Voice activity detection (VAD) separates speech segments from silent segments of an audio signal, and it is valuable for many speech-processing applications because it assists in improving performance and system efficiency; such applications include speech recognition and speaker verification. In this study, K-means, a clustering algorithm, was extended to a thresholding algorithm termed K-means weighted thresholding and was utilized for discriminating voiced/speech segments from silent segments from audio or speech signals. The voice signal was fragmented into frames of 2048 samples, and the spectral power of the frames served as input for computing the threshold value by the extended k-means algorithm; hence, any frame whose spectral power is greater than or equal to the threshold value is considered to part of the voice segments; otherwise, it is tagged as a silent frame. The implemented voice activity detection system achieved outstanding performances with a true acceptance rate (sensitivity), false acceptance rate, true rejection rate (specificity), false rejection rate (miss rate), and a classification accuracy of 100%, 0.025%, 100%, 0%, and 99.97%, respectively.