摘要 |
An objective of the present invention is to provide a new technique of noise alleviation of a model base in voice recognition. In noise correction of a model base, the present invention generates a probability model which is represented as the product of the probability distribution of a mismatch vector (g) (or a clean voice (x)) with an observed value (y) as a factor and a probability distribution of a mismatch vector (g) (or a clean voice (x)) with a reliability index (β) for each band as a factor, executes MMSE estimation on the probability model, and estimates a clean voice estimate value (x^). As a result, each band has an effect on the result of the MMSE estimate with a degree of contribution corresponding to the size of the reliability thereof. Furthermore, as the SNR of the observed voice increases, the output value thereof shifts toward the observed value, and as a result, the output of the front end is optimized. |