Broad phonetic class segmentation study for Thai automatic speech recognition

An automatic broad class segmentation is an important pre-processing step in speech recognition and other speech applications, for example, the speech transcription task to support the phonetic transcription of speech corpus and pronunciation error detection of phone boundaries in language learning applications. This research is aimed at the improvement of the acoustic parameters for the Thai automatic speech recognition system. We proposed acoustic parameters that capture the characteristics of broad manner class of Thai speech. These acoustic parameters are: 1) spectral center of gravity and short time zero crossing rate to classify the silence feature and the continuant feature; and 2) the energy ratio E[0-400] to E[400-6000] to classify the syllabic feature. The results showed 28.09%, 11.0% and 2.41% error reductions for the continuant, the syllabic and the silence features, respectively, when compared to acoustic parameters used in English. The accuracy of 80.46% was obtained from the speech segmentation task and also introduced a 23.46% error reduction when compared to the baseline HMM-MFCC based broad class segmentation. We also found similar performance for word classification in the CVC context when compared to the baseline HMM-MFCC in word recognition tasks.