Describing the phonetics in the underlying speech attributes for deep and interpretable speaker recognition - Avignon Université Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

Describing the phonetics in the underlying speech attributes for deep and interpretable speaker recognition

Imen Ben-Amor
  • Fonction : Auteur
  • PersonId : 1268349
  • IdHAL : imenba
Jean-François Bonastre

Résumé

Deep neural networks have dominated speaker recognition, with a sharp increase in performance associated with increasingly complex models. This comes at the cost of transparency, which poses serious problems for informed decision making. In response, an intrinsically interpretable scoring approach, BA-LR, was recently presented. This method uses an attributebased bottom-up representation of speech, linked with a transparent scoring scheme. For the sake of explainability, the present work adds an analysis of the nature of the attributes, by selecting and quantifying the contributions of the phonetic variables that describe it. We propose two methods based on statistical and surrogate models, respectively. The results reveal that the speech attributes are each well described by a set of descriptive variables. This allows us to propose the first transparent scoring scheme in speaker recognition, where the weights of the phonetic variables contributing to each decision item are known.
Fichier principal
Vignette du fichier
INTERSPEECH_2023-7-1.pdf (644.74 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04155146 , version 1 (07-07-2023)

Identifiants

  • HAL Id : hal-04155146 , version 1

Citer

Imen Ben-Amor, Jean-François Bonastre, Benjamin O'Brien, Pierre-Michel Bousquet. Describing the phonetics in the underlying speech attributes for deep and interpretable speaker recognition. Interspeech 2023, ISCA, Aug 2023, Dublin, Ireland. ⟨hal-04155146⟩

Collections

UNIV-AVIGNON
69 Consultations
175 Téléchargements

Partager

Gmail Facebook X LinkedIn More