TABLE OF CONTENTS We briefly presented in section 3.3, the principles of a text-to-speech synthesis system. Today, major technological advances in speech synthesis systems involve the concatenation of pre-recorded acoustic units, extracted from a corpus of continuous speech. This is known as corpus-based synthesis.
Applications with a limited vocabulary and with constrained syntactic sentence structures can be based on speech synthesis systems that concatenate words, breath groups or sentence fragments. For these systems the units directly stem from the specification of the application, and the consequence is a rigidity of the system [MOB 01]. Conversely, speech quality is usually very close to that of natural speech: co-articulation effects are relatively limited as they occur at the border of the units [ST 99], [LEN 02], [BLA 01]. In this section, we will focus exclusively on studies concerning speech synthesis with an unlimited vocabulary.
In the framework of unlimited vocabulary synthesis, the initial systems were based on phones (which are the acoustic counterparts of the phonemes, defined in the phonological domain). It was therefore natural to consider the phone as the ideal acoustic unit to fabricate sound material: linguistic efficiency and very small inventory of units. However, [HAR 53] showed by a number of experiments on allophones, that assembling phones yields a disastrous result in terms of intelligibility, because the continuity between phones is a key factor in speech perception (co-articulation effects).
We know that diphones, already mentioned in section 3.5.2, have been introduced in order to...
