Study of unknown‐multiple signal‐source clustering problem using ergodic HMM

Abstract

AbstractThe problem in which the input signal sequence is segmented into multiple signal sources and the signal source is estimated appear in a wide range of problems, such as speech information processing and language processing.In this paper, this kind of problem is called the unknown‐multiple signal source clustering problem, and a solution method is proposed based on the ergodic HMM. In ergodic HMM, the state corresponds to the signal source and the symbol sequence output from the state corresponds to the signal sequence. Then, using the Viterbi decoding and the forward decoding, the segmentation point and the category can simultaneously be estimated. As an application of the problem, the classification of the utterances by multiple speakers is attempted. As a result of the experiment, it is shown that the initial parameter values are important in the ergodic HMM, and the LPC cepstrum with a long‐term window is useful as the feature vector reflecting the speaker individuality.