It is reasonable to expect that relevant feautures for speech recognition should be the features that are well heard. Therefore the successful extraction of such relevant acoustic features must respect proper-ties of the auditory periphery. A spectral respresentation incorporating time-frequency forward masking and the way to find optimized masking paremeter is newly proposed This masked spectral representation is efficiently represented by a quefrency domain parameter. Automatic speech recognition experiments have demonstrated that proposed feature powerfully improves performance in word recognition. This spectral representation simulates a perceived spectrum. It suppress temporally stationary spectral properties,such as the effect of microphone frequency charateristics or the speaker-dependent time-invariant spectral feature. The proposed feature is calculated from a cepstral time sequence using matrix lifter. Smoothing charateristics are a function of the time interval between a masker and a signal. The proposed featue outperformed a conventional cepstrum parameter obtained through MFCC especially in noisy environment.
