Rechard Lengagne, Yasuhiro Komori
Automatic Phoneme Segmentation Using
Continuous Hidden Markov Models
Abstract:This report proposes a method for automatic phoneme segmentation of Japanese using continuous mixture density hidden Markov models
(HMMs). Different kinds of training methods have been performed: word,
phrase (bunsetsu) and sentence models are tested after training using single and multiple male speaker word data. The main experiments of this
study, performed on 2 male speakers' utterances use HMMs trained on
data from 8 other male speakers, and yield an average success rate of 95%
in segmentation within a deviation of 30 ms from the "hand"-determined
boundaries.
The problem of word-spotting using "keyword"-segmentation will also
be discussed.