A DATA-DRIVEN JACOBIAN ADAPTATION METHOD FOR THE NOISY SPEECH RECOGNITION AND APPARATUS THEREOF

摘要

One voice recognition method and device, using one kind with data for basis JA (Jacobian determinant adaptation) method, by using reference HMM (hidden Markov model) parameter setting at raising voice sounds recognition performance, the performance having had. One HMM parameter values estimating unit (202) refers to HMM parameter values by repeatedly estimation estimation one, and so pervious HMM parameter values have a possibility that increasing relative to the one of train voice data according to a Bao Mu-Welch method. One Jacobian matrix value estimate unit (206) estimates a Jacobian matrix value according to reference HMM parameter values and train voice data. First and second noise spectrum estimating unit (204,208) thinks that a noise spectrum of voice data is recognized. One phonetic function approximate unit (210) is by using noise spectrum and the approximate phonetic function of Jacobian matrix value. The HMM parameter values for reevaluating unit (212) reevaluate a HMM parameter values according to phonetic function approximation. One viterbi decoder (214) deciphers voice according to the data of the HMM parameter values reevaluated, exports a final recognition result.