Parameter Optimization for a Coarticulation Model Based on Observation and Simulation (国際ワークショップ Frontiers in Speech and Hearing Research)

概要

論文の詳細を見る
A coarticulation model, namely 'carrier model', has been proposed by Dang et al. to improve the naturalness of synthesized speech sounds using a physiological articulatory model. The form of the carrier model offers a good framework to account for the coarticulation in the planning stage, while its parameters need to be refined in order to improve the performance of the model. In this model, we suppose that there is a typical spatial target for each phoneme in the motor space. The objective of this paper is to refine the parameters of the carrier model and learn the typical articulatory targets by reducing the difference between model simulations and observations. These two tasks were combined in a model-based simulation using an optimization framework. To decompose the complicated problem into a set of subproblems, this study adopted a bilevel optimization strategy to obtain the typical target via the carrier model and the parameters for the carrier model based on the typical target. The learning process was carried out by iterations of these two alterative steps. A direct search method was applied in the low level that consists of the processes from the planned targets to speech sounds. A general evaluation was conducted by combining the refined carrier model and the learned typical targets together with the physiological articulatory model.
社団法人電子情報通信学会の論文
2006-03-20