HMM recognition of Hebrew phonemes

This paper describes a speaker dependent system comprising two independent stages. The first stage performs segmentation of the speech to phoneme (or sub phoneme) sections while the second stage performs phoneme classification (labeling). Several labeling procedures based on hidden Markov model are described. A simple continuous speech adaptive phoneme segmentation algorithm is introduced. In an experimental evaluation of the system 93.86% correct boundary marks were received by the suggested adaptive phoneme segmentation algorithm. Recognition rate was 72.38% received for phonemes segmented manually and 68.94% for phonemes segmented using the suggested algorithm. These results are comparable with reported phoneme recognition systems in English and other languages