A study of HMM-based automatic segmentations for Thai continuous speech recognition system

Speech segmentations have been widely using in many speech applications. In speech synthesis, the quality of produced speech depends on the accuracy of labeled acoustic inventory. In speech recognition, segmented utterances according to the labels are usually used as a starting point for training speech models. The segmentation is often manually encoded which is timeconsumption process and has low precision and consistency in some parts of speech. Therefore, the new better technique for speech segmentation using HMM is proposed. In this framework, the effects of manual and automatic segmentation are examined by using the output of final application, the word accuracy of speech recognition. From the experiment, manual segmentation has only 0.59 % better than mono-phone automatic segmentation. The result convinces that tri-phone automatic segmentation will give a better result in the future work.