| 摘要 |
PROBLEM TO BE SOLVED: To solve the problem wherein although the performance of a voice recognition device deteriorates significantly in the circumstances in which there exists long reverberation, which is generally known, and most of the conventional reverberation removal methods require a large amount of calculation is not large, or for those where the amount of calculation is not large, some kind of previous knowledge (reverberation time of a room, etc.) is required. SOLUTION: The coefficient determination in the conventional techniques, in which the multiple value of the coefficient of power spectrum of the past frame is subtracted from the power spectrum of the current frame is calculated at low cost, without having to use the information that incurs calculation cost, such as acoustic model or multi-channel input. As a specific method, a voice power track that properly follows the frame of large power and follows the frame of small power late is obtained, and the interval of which the voice power differs significantly from the voice power of the current frame that is smoothed in the time direction is deduced as being an utterance terminal reverberation interval, and the filter coefficient is decided, in such a manner as to minimize the weighted total sum of the residual voice power in the interval and the subtracted power in the utterance interval (not including the reverberation interval). COPYRIGHT: (C)2008,JPO&INPIT
|
