Modeling the temporal evolution of the vocal tract shape with deep learning - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2023

Modeling the temporal evolution of the vocal tract shape with deep learning

Abstract

This paper overviews our work on the links between coarticulation modeling, approached from the point of view of predicting the vocal tract shape from the phonetic sequence, and the available real-time MRI corpora. Real-time MRI has revolutionized the acquisition of articulatory data through the image quality, the possibility of acquiring and denoising the speech signal, and the possibility of recording corpora containing several thousand sentences. Coarticulation modeling is only possible with the ability to reliably track articulator contours in many images. Tracking techniques using neural networks have provided efficient solutions comparable in reliability to humans. Finally, we show that even if recurrent neural networks trained on these corpora can successfully predict the shape of the vocal tract, it is still necessary to use constraints directly from phonetics to ensure consistency in the prediction.
Fichier principal
Vignette du fichier
ICPHS2023.pdf (10.69 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04209848 , version 1 (18-09-2023)

Licence

Identifiers

  • HAL Id : hal-04209848 , version 1

Cite

Yves Laprie, Vinicius Ribeiro, Karina Isaeva, Justine Leclere, Jacques Felblinger, et al.. Modeling the temporal evolution of the vocal tract shape with deep learning. 20th International Congress on Phonetic Sciences, Aug 2023, Prague (CZ), Czech Republic. ⟨hal-04209848⟩
65 View
24 Download

Share

Gmail Mastodon Facebook X LinkedIn More