On gain errors in a linear prediction vocoder using residual energy excitation

Abstract

Many Linear Prediction (LP) vocoder systems use the LP residual energy method for gain matching. Large gain errors occur however when a low frequency formant is in resonance with voiced excitation impulses. In this paper, the errors are evaluated and their fundamental causes are discussed. It is demonstrated that large errors are due to excessive spectral line suppression of the LP error signal, and to discrepancy between analysis windowing and continuous excitation in the synthesis. Finally, an existing improvement of the method is discussed in the context of the statements.