Phoneme Classification Study for Thai Segment-Based Acoustic Models

This paper presents a study of a phoneme classification experiment in Thai language which is a part of our development of a segment-based speech recognition system. To make feature vectors capture segmental acoustic properties, the phoneme associated with a speech segment is represented using MFCCs extracted from different portions of that segment as well as its duration. The acoustic scoring in our probabilistic framework is composed of finding the probability of a segment belonging to one of a number of phoneme groups and the probability of that segment being a specific phoneme, given the grouping. Phoneme groups are categorized based on manners of articulation and confusions from classification results. Classification using linear discriminant analysis (LDA) with the use of prior probability yields the highest group classification accuracy of 88.5% . When the phoneme group is correctly chosen, the mean probability of a token belonging to that group obtained via LDA is around 0.8. The mean probability is clearly lower when it is incorrectly chosen. This shows that our acoustic scoring of phoneme group has a desirable property