| 摘要 |
PROBLEM TO BE SOLVED: To make a rejection decision with high precision even in case of use in unspecified noisy environment. SOLUTION: A recognized vocabulary collation part 104 calculates the collation likelihood (registered vocabulary likelihood) of a vocabulary to be recognized by using a time series of feature quantities of an input voice and an environment adaptive noise model collation part 106 calculates the collation likelihood (environment noise likelihood) of a noise model adapted to the noisy environment. A rejection decision part 107 compares the registered vocabulary likelihood and environment noise likelihood to decide whether the input voice is noise. When it is decided that the input voice is noise, a noise model adaptation part 108 adaptively updates an environment adaptation noise model by using the input voice. Consequently, the environment adaptation noise model complies with the actual environment and a rejection decision on a noise input can be made with high precision. COPYRIGHT: (C)2003,JPO |
