摘要 |
PROBLEM TO BE SOLVED: To learn stable HMM(hidden Markov model) parameters considering data other than learning data with an operation quantity almost equal to that by a conventional BW method. SOLUTION: In a step S101, plural speech learning data belonging to a category K are read, this is analyzed, and a feature vector (learning data) Yakn is obtained. In a step S102, HMM parametersΘa is estimated using this learning data Yakn. In a step 3, learning data Yakn is discriminated by HMM parametersΘa estimated by the step S102, and erroneous discriminated data Ybkn, in which a category is erroneously discriminated, is found. In a step 4, HMM parametersΘb is estimated using erroneous discriminated data Ybkn found by the step S103. In a step 5, discrimination processing is performed using jointly the HMM parametersΘa estimated by the step 2 and the HMM parametersΘb estimated by the step 103.
|