UNFIXED-LENGTH SOUND MODEL GENERATION DEVICE AND SPEECH RECOGNITION DEVICE

PURPOSE: To provide the unfixed-length sound model generation device which automatically generate an unfixed-length sound model without limiting the model unit of HMM to only a phenome and to provide the speech recognition device which uses it. CONSTITUTION: Phoneme hidden Markov models of plural phonemes are generated on the basis of sound data on a 1st spoken speech sentence of a specific speaker, and some of the phoneme hidden Markov models of plural phonemes are connected on the basis of text data on a 2nd spoken speech sentence different from the 1st spoken speech sentence to generate plural unfixed-length sound models as respective connected hidden Markov models (30). Here, the length of the unfixed-length sound model is determined corresponding to the unfixed- length sound model in the text data so that the frequency of plural mutually adjacent phonemes become maximum, and an unfixed-length sound model is so selected on the basis of the sound data on the 2nd spoken speech sentence of the specific speaker so that the likelihood is maximum. Further, speech recognition is performed by referring to the unfixed-length sound model.