Hierarchical phoneme recognition by hidden Markov models based on multiple feature integration

A method of hierarchical phoneme recognition which utilises the most selective features for each individual phoneme is reported. Input speech patterns are classified into broad classes on the basis of LPC-derived cepstral data. Then, the speech is further classified to a fine-class level using mel-formant data for vowel models only. Hidden Markov models (HMM) are used at both levels of classification.