Investigation on Discriminative Maximum a Posteriori Linear Regression for Speaker Adaptation

In order to increase the discriminative capability of the adapted acoustic model, the maximum mutual information based discriminative maximum a posteriori linear regression (MMI-DMAPLR) adaptation method was proposed. Combining the maximum mutual information criterion with maximum a posteriori (MAP) criterion, a new objective function was designed to estimate the transform parameters of adaptation method based on the linear transformation, to increase the discriminative capability in maximum a posteriori estimation. The experimental results in large vocabulary continuous recognition show that the proposed method can both enhance the match degree between the acoustic model and the test data and the discriminative power of acoustic model. Compared with maximum a posteriori linear regression (MAPLR), the proposed method can obtain 4.8% relative reduction in word error rate when the amount of data is limited.