Linear Prediction

Abstract

This chapter is concerned with the estimation of the spectral envelope of speech signals and its parametric representation. By far the most successful technique, known as linear predictive (LP) analysis, is based on autoregressive (AR) modeling. The application of LP techniques in speech coding is quite natural, as (simplified) models of the vocal tract correspond to AR filters. The underlying algorithmic task of LP modeling is to solve a set of linear equations. Fast and efficient algorithms, such as the Levinson–Durbin algorithm, are available and will be explained in detail. Speech signals can be considered as stationary only for relatively short time intervals between 20 and 400 ms; thus, the coefficients of the model filter change quickly. Therefore, it is advisable to optimize the predictor coefficients frequently. The chapter distinguishes between block–oriented and sequential methods.

Authors 2

1. Peter Vary Aachen Institute of Communication Systems

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Institute of Communication Systems, Aachen, Germany

2. Rainer Martin

   Ruhr University Bochum

   Affiliation as printed

   Ruhr-Universität Bochum, Institute of Communication Acoustics, Bochum, Germany