Context-independent acoustic models for Thai speech recognition

The paper investigates and constructs suitable acoustic models for Thai speech recognition. Many methods are presented, including single phone modeling, base phone modeling, and tone modeling. Context-independent models and gender-independent models are often used in acoustic modeling. Many modeling techniques are given from linguistic-phonetic knowledge of the Thai language. The word. recognition performance of each method is evaluated experimentally using word utterances as training and testing data. The experiments measure the word accuracy of each modeling method when a number of mixtures are varied. The best accuracy is 87.03% from the single phone models without tone modeling (31 models) and 82.03% from the base phone models with tone modeling on vowels (171 models).