On the use of bandpass liftering in speech recognition

In a template-based speech recognition system, distortion measures that compute the distance or dissimilarity between two spectral representations have a strong influence on the performance of the recognizer. Accordingly, extensive comparative studies have been conducted to determine good distortion measures for improved recognition accuracy. Previous studies have shown that the log likelihood ratio measure, the likelihood ratio measure, and the truncated cepstral measures all gave good recognition performance (comparable accuracy) for isolated word recognition tasks. In this paper we extend the interpretation of distortion measures, based upon the observation that measurements of speech spectral envelopes (as normally obtained from standard analysis procedures such as LPC or filter banks) are prone to statistical variations due to window position fluctuations, excitation interference, measurement noise, etc., and may not accurately characterize the true speech spectrum because of analysis model constraints. We have found that these undesirable spectral measurement variations can be partially controlled (i.e., reduced in the level of variation) by appropriate signal processing techniques. In particular, we have found that a bandpass "liftering" process reduces the variability of the statistical components of LPC-based spectral measurements and hence it is desirable to use such a liftering process in a speech recognizer. We have applied this liftering process to several speech recognition tasks: in particular, single frame vowel recognition and isolated word recognition. Using the liftering process, we have been able to achieve an average digit error rate of 1 percent in a speaker-independent isolated digit test. This error rate is about one-half that obtained without the liftering process.