| 摘要 |
PURPOSE: To reduce the amount of overall compute and to increase the processing speed of voice recognition by computing the output probability of a hidden Markov model(HMM) by a probability computing section based on the output probability read out of a probability storage section. CONSTITUTION: A code book 8 which stores specimen vectors that are generated by time dividing specimen voices and a probability storage section 9, in which output probability for every state transition is stored corresponding to the specimen vectors, are beforehand formed. Then, inputted voices are time divided by a vector computing section 4, input vectors are successively generated and specimen vectors which are close to the generated input vectors are selected from the book 8 by a code selecting section 5. Plural output probabilities corresponding to the selected specimen vectors are read out of a probability storage section 9 by a probability computing section 6. The output probability of the HMM is computed by the section 6 based on the read output probabilities. Accordingly, during the computation of the HMM output probabilities, the amount of overall compute is reduced because the output probabilities are not computed from input vectors. |
