Source information representations for synthetic speech: group delay, event and harmonic structures

摘要

A set of requirements on source information representations for a viery higy quality speech resynthesis is introduced. The investigations were motiveted by the fact that resynthesized speech quality by a speech analysis-modification-resynthesis system STRAIGHT has a significant speaker dependency, suggesting that the major contributing factor is the lack of proper synthetics source model. Trials to improve the source model for STRAIGHT have involved introduction of group delay parameters, C/N, P/N and HNR and gave partial success. However, the efforts were not well organized due to the lack of a systematic understanding of these divergent representations. This report introduces a systematic view of these parameters with some illustrative analysis results of natural speech and simulations based on the view.