Fuzzy expectation-maximisation algorithm for speech and speaker recognition
Dat Thanh Tran, Hermann Wagner · 2003
A fuzzy c-means approach to the expectation-maximisation (EM) algorithm is proposed. A family of fuzzy EM algorithms of various degrees of fuzziness is presented, where the EM algorithm is referred to as a fuzzy EM algorithm with degree of fuzziness of one. This fuzzy approach can be applied to EM-style algorithms such as the Baum-Welch algorithm for hidden Markov models, and the EM algorithm for Gaussian mixture models in speech and speaker recognition. The fuzzy EM algorithm for Gaussian mixture models is considered in detail as a demonstration for applying the fuzzy EM algorithm.