Unsupervised Speaker Segmentation Framework Based on Sparse Correlation Feature

With the increasing stress in working and studying, mental health becomes a major problem in the current social research. Generally, researchers can analyze psychological health states by using social perception behavior. The speech signal is an important research direction in this domain. It objectively assesses the mental health of social groups through the extraction and fusion of speech features. Thus, this requires an efficient speech segmentation algorithm. In this paper, we present a new framework of speech segmentation algorithm based on the hybrid of sparse correlation feature with Hidden Markov Model (HMM) as well as Kullback-Leibler Divergence (KLD)while it has been proven to gain higher accuracy. Specifically, HMM method can be used to gain the initial wearer's voice data. Experimental tests and comparisons with different segmentation methods have been conducted to verify the efficacy of the proposed unsupervised method. Very promising results have been obtained.