A discriminative segmental feature transform method is proposed to promote the stability of the frame based method.The feature transform is considered as the sparse high dimensional approximation problem.Firstly

a set of feature transform matrices are estimated by tied-state based training of RDLT (Region Dependent Linear Transform) and m-fMPE (mean-offset feature Minimum Phone Error)

and the transform matrices are integrated into an over-complete dictionary.Then

the speech signal is segmented through force alignment.Finally

following the matching pursuit to optimize the likelihood objective function iteratively

the transform matrices of each segment are selected from the dictionary and the corresponding coefficients are automatic determined in the optimization process.Further

to guarantee the stability of the transform matrices

a correlation measurement is introduced to remove the correlated basis in the recurrence process.The experimental results show that

compared with the traditional RDLT method

when the acoustic model is trained with maximum likelihood and discriminative training criterion separately

the recognition performance can be improved by 1.63% and 2.23% respectively.The method can also be applied to speech enhancement and model discriminative training.