Modified Filterbank Analysis Features for Speech Recognition
Deividas Eringis · 2015
Auditory model based feature systems include filterbank analysis and nonlinear compression of the speech signals. The Mel Frequency Cepstral Coefficients (MFCC) is the state- of-the-art feature system employing this auditory model. In this paper we proposed to modify MFCC analysis by applying power nonlinearity operator instead of logarithmic and to modify the size of filterbank. Power nonlinear operator caused increased recognition rate of deteriorated speech by 2.4 %. In combination with reduced filterbank size (down to 20 filters) power nonlinearity enhanced robustness of speech recognition: the gain of recognition rate varied from 0.6 % to 3 % in comparison with common MFCC features for different noise levels.