Multilingual Models with Language Embeddings for Low-resource Speech Recognition - Université Paris-Saclay Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Multilingual Models with Language Embeddings for Low-resource Speech Recognition

Résumé

Speech recognition for low-resource languages remains challenging and can be addressed with techniques such as multilingual modeling and transfer learning. In this work, we explore several solutions to the multilingual training problem: training monolingual models with multilingual features, adapting a multilingual model with transfer learning and using language embeddings as additional features. To develop practical solutions we focus our work on medium size hybrid ASR models. The multilingual models are trained on 270 hours of iARPA Babel data from 25 languages, and results are reported on 4 Babel languages for the Limited Language Pack (LLP) condition. The results show that adapting a multilingual acoustic model with language embeddings is an effective solution, outperforming the baseline monolingual models, and providing comparable results to models based on state-of-the-art XLSR-53 features but with the advantage of needing 15 times fewer parameters.
Fichier principal
Vignette du fichier
lamyeemui23_sigul.pdf (201.12 Ko) Télécharger le fichier
Origine : Fichiers éditeurs autorisés sur une archive ouverte

Dates et versions

hal-04397807 , version 1 (16-01-2024)

Identifiants

Citer

Léa-Marie Lam-Yee-Mui, Waad Ben Kheder, Viet-Bac Le, Claude Barras, Jean-Luc Gauvain. Multilingual Models with Language Embeddings for Low-resource Speech Recognition. 2nd Annual Meeting of the ELRA/ISCA SIG on Under-resourced Languages (SIGUL 2023), ELRA/ISCA, Aug 2023, Dublin, Ireland. pp.83-87, ⟨10.21437/SIGUL.2023-18⟩. ⟨hal-04397807⟩
22 Consultations
20 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More