Skip to Main content Skip to Navigation
Master Thesis

Dictée automatique de textes à haute voix : analyse de corpus et modélisation des stratégies d'énonciation

Abstract : The work presented in this paper is in the context of Text-To-Speech synthesis (TTS). This work aims to develop an automatic dictation system that converts written text into spoken words, for use by CM1 and CM2 students (4th and 5th grade in French primary school). This system includes four modules : 1) a morpho-syntactic analysis module which extracts tree data structure from orthographic strings; 2) a prosodic segmentation module which divides the utterances into prosodic groups and relates them to their propositional content; 3) a prosody generation module which calculates prosody automatically (fundamental frequency contour and speech rhythm), to be applied to those groups based on phonotactic features; 4) a synthesis module which converts this segmental and supra-segmental data into an acoustic signal. The syntactic parsing is first performed. The syntactic tree is then further projected onto the syntagmatic axis : sentence-internal prosodic markers are supposed to cue the dependency relations between adjacent constituents. Prosody generation is then performed by associating these markers with multiparametric contours via so-called contour generators. These contour generators are implemented as feed-forward neural networks and trained thanks to an iterative analysis-by-synthesis process (see description of the SFC "Superposition of Functional Contours" model in Bailly and Holm (2003) using a corpus of dictations. We finally assessed the quality of the modeling with a synthesis system available at GIPSA-Lab. We showed that strategy-specific contour generators are able to capture the slow rate of speech flow, the increased rate of word and phrase segmentation as well as changes in the melodic contours of the four versions for each phrase.
Document type :
Master Thesis
Complete list of metadatas

Cited literature [10 references]  Display  Hide  Download

https://dumas.ccsd.cnrs.fr/dumas-01062033
Contributor : Uga - Bulles <>
Submitted on : Tuesday, September 9, 2014 - 10:08:59 AM
Last modification on : Wednesday, July 15, 2020 - 1:06:02 PM
Long-term archiving on: : Wednesday, December 10, 2014 - 11:10:30 AM

Identifiers

  • HAL Id : dumas-01062033, version 1

Citation

Mejdi Slim. Dictée automatique de textes à haute voix : analyse de corpus et modélisation des stratégies d'énonciation. Sciences de l'Homme et Société. 2014. ⟨dumas-01062033⟩

Share

Metrics

Record views

1057

Files downloads

1715