VOICE RECOGNITION DEVICE

PURPOSE: To improve a voice recognition rate without increasing a learning sample quantity for voice recognition by merging and then re-learning a phonemic environment judged to be most similar based on a likelihood as a reference outputted from an unstational signal source. CONSTITUTION: A voice signal input part 11 converts voice signals into digital values. A feature extraction part 12 computes featured values from the input signals for every frame. A data storage part 13 stores HMM as a learned unstational signal source having three kinds of phonemic environment factors of precedence, pertinence and succedence. A recognition result decision part 14 computes an output probability of the input phoneme and also decides the recognition result by two-stage DP algorithm. In HMM prepared for each phonemic environment space in which a preceding phoneme and a succeeding phoneme existing in learning samples are considered, the likelihood for a learning sample set in the individual phonemic environment spaces is obtained for each HMM, the phonemic environment space with a least difference in likelihood is judged to be the most similar space and merged for relearning.