Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Sparse Vector Linear Prediction Matrices with Multidiagonal Structure

Davor Petrinovic, Davorka Petrinovic

Faculty of Electrical and Computer Engineering, Department of Electronic Systems and Information Processing, University of Zagreb, Croatia

A modification of the classical vector linear prediction (VLP) problem is presented. The introduced technique called the sparse VLP (sVLP) is based on the assumption that each component of a single LSF vector is highly correlated only to a few neighboring vector components of consecutive vectors, while the correlation between distant vector components can be ignored. This leads to simplification of predictor matrices in a way that for a chosen number of neighboring components, predictor matrices obtain multidiagonal form. It is shown that prediction gain resulting from sVLP is only slightly lower than for the case of full matrix predictors but with significant reduction of computation, both for coding and predictor design.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  Petrinovic, Davor / Petrinovic, Davorka (1999): "Sparse vector linear prediction matrices with multidiagonal structure", In EUROSPEECH'99, 1483-1486.