Promises and Limitations of Self-supervised Learning for Automatic Speech Processing - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

Promises and Limitations of Self-supervised Learning for Automatic Speech Processing

Promesses et Limites de l’Apprentissage Auto-Supervisé pour le Traitement Automatique de la Parole

Résumé

Self-supervised learning (SSL) has recently been successfully introduced as a training strategy for Transformerbased neural models. Thanks to this approach, these models are now able to construct speech representations by using only audio data, without any manual labels (i.e. no supervision). Once trained, they can be leveraged for training competitive endto-end models for speech processing with smaller amounts of annotated data. Moreover, when the available annotated data is plenty, automatic speech recognition (ASR) and translation (AST) systems based on these SSL models are now the new state of the art. In this work, we are interested in their application in challenging settings that are relevant for security. We measure the robustness of a French-based SSL model to African accent, and we present some promising but limited results for speech translation without the use of transcriptions.
Fichier principal
Vignette du fichier
92.pdf (464.13 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03881745 , version 1 (02-12-2022)

Licence

Paternité - Pas d'utilisation commerciale - Pas de modification

Identifiants

  • HAL Id : hal-03881745 , version 1

Citer

Lucas Maison, Marcely Zanon Boito, Yannick Estève. Promises and Limitations of Self-supervised Learning for Automatic Speech Processing. Conference on Artificial Intelligence for Defense, DGA Maîtrise de l'Information, Nov 2022, Rennes, France. ⟨hal-03881745⟩
59 Consultations
158 Téléchargements

Partager

Gmail Facebook X LinkedIn More