A study of various linguistic effects on tone recognition in Thai continuous

Abstract

To study the various linguistic effects, i.e., syllable structure, coarticulation, intonation and stress on tone recognition in Thai continuous speech. Tone models for compensating these effects are also developed. We first study the effect of initial consonants, vowels, and final consonants on tone recognition in isolation. Based on the observation on Fo contours, we proposed a novel tone feature set. The new feature set achieved better recognition rates than the conventional tone feature sets. We also explored several combinations of classifier schemes and found that the combinations of classifiers were superior to a single classifier. Next, we developed a basic tone recognition framework for Thai continuous speech. The framework consisted of tone models used to parameterize Fo contours of tones and a classifier used to evaluate the performance of the tone models. We conducted experiments to construct the tone models by concentrating on tone features, frequency scales, normalization techniques, and tone critical segments. The classifier was developed using a feed-forward neural network. Next, we focussed on tone coarticulation effect. We have proposed a feature set called "contextual tone features" that captured the Fo realizations of the neighboring syllables. The features provided the best tone error reduction rates of 56.17%, 42.47%, and 42.42% for Thai Proverb Corpus (TPC), Potisuk-1999 Corpus (PC-99), and Thai Anumal Story Corpus (TASC), respectively. Furthermore, we explored the context-dependent tone model (CD-T-175) and developed a novel model, half-tone model (H-T-30). Both models increased recognition rates, but the training time of H-T-30 was one-fourth of CD-T-175. Next, we studied the effect of intonation on tone recognition. We obtained two methods, i.e., beginning-point intonation normalization and center-point intonation normalization methods to compensate the intonation effect. Both methods significantly increased recognition rates. The best error reduction rates of 22.20% and 16.84% were achieved for TASC and TPC, respectively. Next, we concentrated on stress effect. We first performed two empirical experiments of stress detection on pairs of ambiguous words and poly-syllabic words. We explored acoustic features, i.e., duration, energy, and Fo extracted from several linguistic units, i.e., vowel, syllable and rhyme units. The rhyme unit outperformed the other units for stress detection. We then performed an empirical study of tone recognition. We have proposed two methods, i.e., separated stress method (SSM) and incorporated stress feature method (ISFM). Both methods increased the tone recognition rates. We additionally incorporated ISFMs into the tone model and found that TSFM improved the recognition rates. The highest error reduction rates of 32.43% and 27.16% were reported for TPC and TASC, respectively. Finally, we integrated several refined tone models into a syllable-based speech recognition system to enhance the recognition performance. We achieved the best error reduction rates of 85.16% and 75.06% for TPC and TASC, respectively.