TRIPHONE BASED CONTINUOUS SPEECH RECOGNITION SYSTEM FOR TURKISH LANGUAGE USING HIDDEN MARKOV MODEL

摘要

This paper introduces a system which is designed tornperform a relatively accurate transcription of speech andrnin particular, continuous speech recognition based onrntriphone model for Turkish language. Turkish is generallyrndifferent from Indo-European languages (English,rnSpanish, French, German etc.) by its agglutinative andrnsuffixing morphology. Therefore vocabulary growth raternis very high and as a consequence, constructing arncontinuous speech recognition system for Turkish basedrnon whole words is not feasible. By considering this fact inrnthis paper, acoustic models which are based on triphones,rnare modelled as five state Hidden Markov Modelsrn(HMM). Mel-Frequency Cepstral Coefficients (MFCC)rnapproach was preferred as the feature vector extractionrnmethod and training is done using embedding training thatrnuses Baum-Welch re-estimation. Recognition isrnimplemented on a search network which can be ultimatelyrnseen as HMM states connected by transitions and ViterbirnToken Passing algorithm runs on this network to find thernmostly likely state sequence according to the utterance.rnAlso to make a more accurate recognition bigramrnlanguage model is constructed.