Kiyoaki Aikawa, Hideki Kawahara, Yoh'ichi Tohkura
Dynamic Cepstrum Parameter Incorporating Time-Frequency
Masking and Its Application Speech Recognition
Abstract:A Dynamic Cepstrum parameter is proposed that incorporates the time-frequency characteristics of forward masking. Psychological research reports that the forward masking
pattern becomes more wide-spread over the frequency axis as the masker-signal time-interval
increases. To simulate the masking characteristics, a novel lifter array operation
is derived. The Dynamic Cepstrum can represent both the instantaneous and transitional
aspects of speech spectra. The proposed parameter is superior to the conventional delta
cepstrum in extracting high temporal resolution spectral dynamics, and outperforms the
conventional cepstrum in phoneme recognition.
This technical report is the detailed version of the paper presented at the 124th Meeting
of the Acoustical Society of America (J. Acoust. Soc. Am., Vol. 92, No.4, Pt. 2,
pp.2476, 5pSP5, Oct. 1992).