An LVQ2-trained connected-phoneme hidden Markov model for automatic phonetic segmentation and labeling

• DOI: 10.1121/1.406215
• Publication Date: 2005-10-13T21:47:23Z
• Authors: Y. Chun Kuo; Janet C. Rutledge
• Keywords: Viterbi algorithm; Sequence labeling

A connected-phoneme hidden Markov model (HMM) is proposed to perform automatic segmentation and labeling. Individual phonetic models are first created by a left-to-right HMM. The large HMM formed grouping all these together. Therefore, each state of this big uniquely represents an English phoneme. not trained the Viterbi algorithm since most probable sequence dose necessarily yield correct Learning vector quantization (LVQ2) used train such that phoneme confusions can be reduced. has two potential advantages over existing speech recognition schemes. (1) With aid unique representation HMM, more insight into characteristics gained, which essential for improvement recognizers. Errors caused insertion, deletion, substitution properly analyzed adjusted. (2) computation load LVQ2 training considerably less than training. also suitable limited data base.