| 摘要 |
PURPOSE:To improve the recognition speed of a voice recognition device employing a hidden Markov model(HMM). CONSTITUTION:A voice dictionary file 5 stores information including the number of states, a state transition probability and the output probability of feature data (level) of the learned voice model relative to the work voice segmented from a beforehand inputted uttered voice. A work voice segmenting section 2 segments a word voice from the recognition object voice which is uttered and inputted from a voice inputted means 1. A feature extracting section 3 extracts feature data from the word voice. A state number estimating section 4 estimates the number of states of a recognition object voice. A prediction discriminating section 6 preliminary selects voice models, in which are number of states in arrange between the estimated state number and a prescribed region, from the voice models of the file 5. A collation deciding section 7 computes the liklihood of the preliminary selected voice models and the recognition object and performs a recognition decision. A decision result output section 7 outputs the discrimination result. |
