Skip to Main content Skip to Navigation
Theses

Segmentation automatique de parole en phones. Correction d'étiquetage par l'introduction de mesures de confiance

Samir Nefti 1
1 CORDIAL - Human-machine spoken dialogue
IRISA - Institut de Recherche en Informatique et Systèmes Aléatoires, INRIA Rennes, ENSSAT - École Nationale Supérieure des Sciences Appliquées et de Technologie
Abstract : A speech synthesis system which operates by concatenation of acoustic units, needs to a database of these units, which is built starting from a single speaker speech corpus segmented in acoustic elements, generally phonetic. To reach a sufficient synthetic speech quality, this database must be richly provided, and consequently requires a corpus of several hours of speech.
The manual segmentation of such a corpus is tedious, and from there occurs the interest of the automatic segmentation. The automatic methods produce a segmentation of quality approximately equivalent to that of a manual segmentation, but provided that the exact phonetic transcriptions of the utterances are available. However, the manual phonetic transcription of the speech corpus is also tedious.
This study concerns the automatic segmentation of speech into phones, which uses phonetic transcriptions automatically produced from text. It relates to the detection and the correction of the labeling errors which occur generally in these automatic phonetic transcriptions. The results obtained in this study are significantly positive.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00122091
Contributor : Samir Nefti <>
Submitted on : Tuesday, December 26, 2006 - 1:14:38 PM
Last modification on : Friday, July 10, 2020 - 4:20:30 PM
Long-term archiving on: : Friday, September 21, 2012 - 9:46:37 AM

Identifiers

  • HAL Id : tel-00122091, version 1

Citation

Samir Nefti. Segmentation automatique de parole en phones. Correction d'étiquetage par l'introduction de mesures de confiance. Interface homme-machine [cs.HC]. Université Rennes 1, 2004. Français. ⟨tel-00122091⟩

Share

Metrics

Record views

653

Files downloads

1269