Like human recognition performance, computer recognition performance degrades inevitably in the presence of interference and distortion. There are two basic approaches to minimizing the degradation: (1) seek new acoustic features that yield performance that is less sensitive to signal corruption; (2) try to "see through" the corruption and to the "true" values of the information-bearing features. We suggest an approach applying information theory to estimate the underlying acoustic features. The proposed work will extend previous results in which Minimum Relative-Entropy (MRE) techniques were used to provide robust estimation of LPC parameters in the presence of helicopter noise. Analogous MRE techniques will be used to develop robust estimators for other "traditional" acoustic features (e.g., cepstral coefficients and line spectrum frequencies). Furthermore, we will build on recent results that have provided an information-theoretic description of the cochlear representation of speech. Certain cochlear models have been shown to have a certain degree of intrinsic noise-immunity. We hope to obtain additional noise-immunity using the MRE approach.Anticipated benefits/potential applications:There is strong empirical evidence that out approach will lead to improved recognition performance. Further, the work will help to extend a general theory that should be useful beyond the estimation of the acoustic proposed for study in this project. On-going applications are likely through several avenues: via Entropic's new role as a technology-transfer agent for the DARPA Speech and Language Systems (SLS) program; another is through Entropic's commercial signal processing software packages ESPS and waves+.
