Possibilities of Utilizing Linguistic Information in Development of Advanced Speech Signal Compression Technologies

Speech waveforms are known to be remarkably difficult to compress. Lossless entropy compression methods yield about 30% reduction of the initial waveform size. Delta-compression (and also somewhat similar to it methods based on the ideas suggested by A. Haar and I. Daubechies) allow to improve this result by another 30%. Lossy compression methods yielding much higher compression rates take advantage of the psychoacoustic peculiarities of sound perception by removing certain components of the speech signal that have the smallest effect over its perceived quality, e.g. portions of the spectrum that are “in the shadow” of prominent spectral peaks. It must be observed that all the abovementioned ideas are based on the existence of some a priori knowledge about such kind of data as the speech signal (for instance, that the ordinates of two adjacent points of the speech waveform normally differ by a small value, or that the spectrum of the speech signal usually contains several prominent peaks that determine the perceived quality of the sound). Following this logic, it can be suggested that the a priori knowledge that the speech signal is a container enclosing linguistic data can just as well be used to significantly improve the efficiency of both lossy and lossless compression technologies by including in the data source model the regularities appertaining to the linguistic component of the speech signal. This paper deals with the possible results of this suggestion being tested.