Towards an efficient computation of masks for multichannel speech enhancement - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2022

Towards an efficient computation of masks for multichannel speech enhancement

Résumé

Most of recent advances in speech enhancement (SE) have been enabled by the use of complex deep neural network (DNN) architectures. Although these results are convincing, they are not yet applicable in small wearable devices like hearing aids. In this paper, we propose a DNN-based SE which benefits from the spatial information to simplify the requirements of the DNN architecture. We show that the DNN inference is the most time and energy consuming step and we simplify the architecture of a convolutional recurrent neural network by removing its recurrent layer. This achieves comparable performance to the initial architecture, while reducing the processing time and energy consumption by a factor of 4.4.
Fichier principal
Vignette du fichier
paper_eusipco.pdf (288.63 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03604983 , version 1 (10-03-2022)

Identifiants

  • HAL Id : hal-03604983 , version 1

Citer

Louis Delebecque, Romain Serizel, Nicolas Furnon. Towards an efficient computation of masks for multichannel speech enhancement. 2022. ⟨hal-03604983⟩
194 Consultations
209 Téléchargements

Partager

Gmail Facebook X LinkedIn More