Normalization of Spectral Information in a Pulse-Excited Channel-Vocoder System

Abstract

Speech synthesis in vocoder systems works under the assumption that the two constituents of the speech description in such systems are independent of each other. As this assumption is not completely true, the synthesized speech signals show distortions that can be avoided by proper preprocessing of the transmitted spectrum and pitch information before using it to control the synthetizer. The already known “spectrum-flattening” technique modifies the pitch-control signals. A new method is proposed that tries to secure “orthogonality” of the spectrum description and the pitch description by compensating pitch influence on the spectral description. It offers several advantages, especially in speech-processing installations where a channel vocoder is utilized in cooperation with a digital computer. The results of intelligibility measurements for speech synthetized from unprocessed spectrum data as well as from data processed by a computer are reported.