Speech Data Rate Reduction Part II: Applicability of Sensitivity and Error Analysis

Abstract

This paper deals with the application of modern estimation techniques to the problem of speech data rate reduction. It is desirable to adaptively identify and quantitize the parameters of the speech model. These paramaters cannot be identified and quantized exactly; the performance of the predictor is thereby degraded and this could prevent data reduction. In many cases it is desirable to emply a suboptimal predictor in order to simplify the algorithms, and predictor performance is again degraded. This paper develops sensitivity and error analysis as a potential method for determining quantitatively how speech data reduction system performance is degraded by imprecise parameter knowledge or suboptimal filtering. An intended use of the sensitivity and error analysis algorithms is to determine parameter identification and model structure requirements of configuration concepts for adaptive speech digitizers. First, sensitivity and error analysis algorithms are presented that form the basis for the remainder of the work. The algorithms are then used to determine how imprecise knowledge of vocal tract parameters degrades predictor performance in speech. Transversal filters have previously been proposed for this application. The sensitivity analysis algorithms are then used to determine when and by how much the transverse filter is suboptimal to the Kalman filter. In particular, the question of how effectively a higher order of all-pole model approximates a system with zeros is answered, as this question is of considerable importance in speech. Finally, the physical significance of the innovations process in speech data rate reduction is studied.