Abstract
Currently, the core of Thai speech recognition algorithms focuses on word recognition. However, such algorithms are not appropriate to construct the Speech-to-Text system since the ultimate goal in Speech-to-Text system is to recognize continuous speech states from any speaker of a given language. The meaning in each given language and its sound can be determined by phonemes which are slightly different for each language. The variability in each speaker's voice, for instance, accents, gender and speech style, and the tonal language such as Thai language can create rather different speech signals for the same word. Thus, phoneme recognition is more difficult to perform. Since segmentation takes place prior to recognition in such systems, an incorrect segmentation invariably leads to incorrect recognition results. We proposed a method for phoneme segmentation that based on Discrete Wavelet Transform (DWT). To verify our method, we performed experiments on eleven speakers: five males and six females. Each speaker pronounced one hundred and thirty Thai words. Then, we evaluated the performance of our method by synthesizing the new words from the obtained phonemes. The speech synthesis of the new words was then observed by humans to compare with the natural-sounding speech. The results indicated 95% accuracy.