METHOD AND DEVICE FOR UNREGISTERED WORD DETECTION, AND VOICE RECOGNITION DEVICE

摘要

PROBLEM TO BE SOLVED: To improve the voice recognition performance by finding cumulative values from the conditioned probability of an output in respective states of a word standard model for a voice recognition result and comparing them with a reference value. SOLUTION: A word matching part 15 once receiving a speaking detection signal 13A from a speaking detection part 13 reads each word standard model 16A out of a word standard model storage part 16 and matches it against a delayed acoustic analytic result 12A. Then once an on-hook detection signal 14A is received from an on-hook detection part 14, information on the word standard model having maximum consistency is outputted as a recognition result 15A. Further, an unregistered word detection part 17 finds the conditional probability of the output in the respective states of the word standard model from the time correspondence among the states of the word standard model of the recognition result 15A and compares their cumulative values with a reference value to detect whether or not the input voice is an unregistered word, and sends the result 17A to a recognition result output part 18. When the cumulative value of the conditional probability is less than the reference value, it is decided that the input voice is an unregistered word.