Chinese Dialect Identification Using SC-GMM
Ming Liang Gu, Biao Zhang · Advanced materials research · 2012
Gaussian mixture model (GMM) is a sort of effective identification method in Chinese dialects identification, estimating GMM parameters is always an important step in building a state-of-the-art speech processing system. One of the most widely used approaches is maximum- likelihood estimation, where parameters of class-specific distributions are estimated using Expectation Maximization algorithm(EM). Initial parameters have great influence on the convergence of EM algorithm, so how to initialize GMM parameters is a key problem. In this paper, we apply spectral clustering(SC) to initialize GMM parameters. Experimental results prove that using spectral clustering algorithm to initialize GMM parameters is superior to traditional K-Means method and identification system has a higher recognition rate.