Assignment of segmental duration in text-to-speech synthesis

Research output: Contribution to journalArticle

89 Scopus citations

Abstract

In natural speech, durations of phonetic segments are strongly dependent on contextual factors. For synthetic speech to sound natural, the module for computing segmental duration (the duration system) must mimic these contextual effects as closely as possible. Construction of duration system is obstructed by two facets of segmental duration: (1) interactions between contextual factors, and (2) sparsity of training data. This paper describes a new duration system in which a central role is played by duration models, in the form of equations consisting of sums and products such as in: duration (/i/, voiced, stressed)=A(/i/)+B(voiced)+C(stressed). These models, which w call sums-of-products models, can capture the types of interaction patterns often found in duration data, where one factor typically amplifies - but does not reverse - the effects of other factors. Yet, these models are mathematically sufficiently tractable for robust parameter estimation in the presence of severe sparsity. The overall architecture of the system consists of a category structure, or tree, that divides the space into similar-behaved cases; for each of these categories a separate sums-of-products model is developed and its parameters are estimated. Perceptual evaluation results are reported for a implementation in the AT and T Bell Laboratories text-to-speech system.

Original languageEnglish (US)
Pages (from-to)95-128
Number of pages34
JournalComputer Speech and Language
Volume8
Issue number2
DOIs
StatePublished - Apr 1994

    Fingerprint

ASJC Scopus subject areas

  • Software
  • Theoretical Computer Science
  • Human-Computer Interaction

Cite this