Abstract
This paper proposes a new applied method to compress the VCV (Vowel-Consonant-Vowel) speech spectrum pattern for synthesis by rule using a Genetic Algorithm (GA). In this method synthesized speech is generated by using parameters from a subset of all of the analyzed VCV frame data. The subset which can generate high-quality synthesized speech is selected automatically by using an optimum combination technique based on GA. We regarded “individuals” in GA as the random selected subsets and defined “fitness” as the inverse of the perceptually based weighting error which is extracted from the difference between natural speech and synthesized speech. Furthermore by using a “generation-dependent mutation width”, the convergence speed can be greatly increased as compared to a ordinary mutation technique. The results from a listening experiment indicate that the synthetic speech generated from the compressed spectrum pattern is of sufficient quality to be of practical use.
