ISCA Archive SSW 1990
ISCA Archive SSW 1990

Automatic derivation of segment models for synthesis by rule

Wendy J. Holmes, David J. B. Pearce

This study aims to improve synthesis quality using the Holmes, Mattingly and Shearme (1964) phonetic-level synthesis-by-rule (SbR) method, both by increasing the inventory of allophone segments and by automatically optimizing the values of the segment-table entries. Every occurrence of each phoneme is first optimized using a separate segment model. Initial estimates are iteratively refined using an analysis-by-synthesis procedure based on comparisons between the natural and rule-synthesized speech spectra, so imposing the inherent continuity constraints of the SbR model. The paper describes this automatic process, whose output is a set of individual segment tables for high quality segmental copy synthesis. These individual tables will later be combined to form allophone models for improved synthesis by rule.