Word synthesis based on line spectrum pairs

Abstract

The author described the initial investigation of a synthetic speech system based on line spectrum pair (LSP) analysis of the speech signal. The synthesizer contains a library of stored LSP speech segments extracted from natural speech. These segments are modified as necessary by a small set of context-sensitive rules and then concatenated to generate high-quality natural-sounding speech. Tests of a preliminary system produced MRT and DRT scores of 87.3 and 79.7, respectively. The LSP vocabulary synthesizer is limited to utterances of a single syllable, but future research will expand its capabilities to allow implementation of a full text-to-speech system.<>