Sixth International Conference on Spoken Language Processing
The measurements of speech spectral envelopes may not accurately characterize the true speech spectrum because of analysis model constraints, such as window position fluctuations, excitation interference, and measurement noise. Juang found that these undesirable spectral measurement variations could be partially reduced by bandpass liftering, and gained better results in speech recognition. In this paper, we propose a new bandpass lifiering process for speaker recognition. The new lifiering process reduces the error rate of 43% than those without the liftering and 9% than Juang's method in speaker recognition. from the experimental result, we found that the spectral contours contain both speech and speaker information and the fine spectral structures contain speaker-discriminating information.
Bibliographic reference. Zhen, Bin / Wu, Xihong / Liu, Zhimin / Chi, Huisheng (2000): "On the use of bandpass liftering in speaker recognition", In ICSLP-2000, vol.3, 933-936.