TABLE OF CONTENTS The evaluation of speech synthesis is a research and application domain in direct connection with the current development of synthesis. Indeed, if a method existed to precisely evaluate and diagnose the quality defects of a speech synthesizer, it would be easier to overcome them or at least to find solutions. Therefore, progress made in evaluating and analyzing synthetic voice quality conditions intimately that of the synthesis algorithm themselves.
Let us recall that the term speech synthesis can cover very different types of systems:
restitution of a pre-recorded message;
concatenation of words or syntactic groups, with a fully determined vocabulary and phraseology;
text-to-speech synthesis, i.e. the pronunciation of any text with an unlimited vocabulary; this is a much more complex task and its result is much less pleasant to the ear.
In the first two cases, the evaluation of speech synthesis is necessarily global. Conversely, for text-to-speech synthesis, the evaluation can be both global and specific to certain aspects of the system. Referring to the description of sections 3.4 and 3.5, three functional components can be identified in a text-to-speech synthesis system:
text analysis, for transforming the written text into a sequence of phonetic symbols (which represent the pronunciation of the text) and prosodic symbols (for marking accentuation and intonation). This first component is primarily related to the automatic language-processing field;
the symbolic description of the text must be converted into numerical parameters (pitch contour, pauses, voice strength, etc.) using...
