In this thesis, New technique named two level phoneme classification(TLPC) for the isolated word recognition system is proposed. This technique is especially effective in speaker dependent large vocabulary word recognition system or speaker independent word recognition system. Word templates are represented as sequences of discrete phoneme templates which are determined by TLPC from a training set of word utterances. During the first level of TLPC, word utterances are scanned for vowel detect.
During the second level, input words are segmented into phonemes, and in training procedures, phonemes are clustered for multiple phoneme templates. After TLPC is performed, input utterances are represented as string of phoneme templates and this string is compared with reference word templates.