In order to improve robustness of speech recognition

a new method of feature combination is proposed.Weighted optimization of Mel frequency cepstral coefficients（ MFCC） is carried out based on the F ratio.Meanwhile

several features are extracted for combination in different ways as input of hidden Markov model（ HMM） for training. Then

the best combination with anti-noise property is descended by PCA. Finally

the SVM classifier is added as a post processor. The experiment shows that the combination of improved MFCC

short-time average energy and Teager energy operator performs best

whose recognition rate is 90. 48 %. And dimensionality reduction based on PCA brings a 0. 4 % reduction on recognition rate. But the speed of calculation is promoted.After adding post processor

the total recognition rate of the system reaches 95. 25 %. The recognition efficiency of the system and the ability of system classification and decision-making are improved. Compared with the conventional recognition method

the accuracy rate is improved.