Analysis of data collected in listening tests for the purpose of evaluation of concatenation cost functions

Abstract

In this paper we present an analysis of data, which were collected in listening tests, and are planned to be used for the development and evaluation of concatenation cost functions for unit selection based TTS systems. The aim of the analysis was to evaluate a "richness" of the collected data with respect to the intended utilization. No effort was made to propose a new method for measuring concatenation artifacts. The study was limited to two speakers (male and female), and five short Czech vowels as these sounds are characterized by being highly energetic and having rich spectral content, which induces complexity and wide range of possible discontinuities at concatenation points. © 2011 Springer-Verlag.