Automatic Phoneme Segmentation With Relaxed Textual Constraints - Université Pierre et Marie Curie Accéder directement au contenu
Communication Dans Un Congrès Année : 2008

Automatic Phoneme Segmentation With Relaxed Textual Constraints

Pierre Lanchantin
  • Fonction : Auteur
  • PersonId : 919059
Andrew Cameron Morris
  • Fonction : Auteur
Christophe Veaux
  • Fonction : Auteur
  • PersonId : 919058

Résumé

Very high quality text-to-speech synthesis can be achieved by unit selection in a large recorded speech corpus [1]. This technique uses some optimal choice of speech units (e.g. phones) in the corpus and concatenates them to produce speech output. For various reasons, synthesis sometimes has to be done from existing recordings (rushes) and possibly without a text transcription. But, when possible, the text of the corpus and the speaker are carefully chosen for best phonetic and contextual covering, for good voice quality and pronunciation, and the speaker is recorded in excellent conditions. Good phonetic coverage requires at least 5 hours of speech. Accurate segmentation of the phonetic units in such a large recording is a crucial step for speech synthesis quality. While this can be automated to some extent, it will generally require costly manual correction. This paper presents the development of such an HMM-based phoneme segmentation system designed for corpus construction.
Fichier non déposé

Dates et versions

hal-01161385 , version 1 (05-01-2016)

Identifiants

  • HAL Id : hal-01161385 , version 1

Citer

Pierre Lanchantin, Andrew Cameron Morris, Xavier Rodet, Christophe Veaux. Automatic Phoneme Segmentation With Relaxed Textual Constraints. Language Resources and Evaluation Conference (LREC2008), May 2008, Marrakech, Morocco. pp.1-1. ⟨hal-01161385⟩
95 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More