Automatic Evaluation of Speech Intelligibility Based on i-vectors in the Context of Head and Neck Cancers - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2018

Automatic Evaluation of Speech Intelligibility Based on i-vectors in the Context of Head and Neck Cancers

Imed Laaridh
  • Fonction : Auteur
  • PersonId : 996854
  • IdRef : 223527904
Corinne Fredouille
Alain Ghio
Muriel Lalain

Résumé

In disordered speech context, and despite its well-known sub-jectivity, perceptual evaluation is still the most commonly used method in clinical practice to evaluate the intelligibility level of patients' speech productions. However, and thanks to increasing computing power, automatic speech processing systems have witnessed a democratization in terms of users and application areas including the medical practice. In this paper, we evaluate an automatic approach for the prediction of cancer patients' speech intelligibility based on the representation of the speech acoustics in the total variability subspace based on the i-vector paradigm. Experimental evaluations of the proposed predictive approach have shown a very high correlation rate with perceptual intelligibility when applied on the French speech corpora C2SI (r=0.84). They have also demonstrated the robustness of the approach when using a limited amount of disordered speech per patient, which may lead to the redesign and alleviation of the test protocols usually used in disordered speech evaluation context.
Fichier principal
Vignette du fichier
1266.pdf (477.7 Ko) Télécharger le fichier
Origine : Fichiers éditeurs autorisés sur une archive ouverte
Loading...

Dates et versions

hal-01962170 , version 1 (20-12-2018)

Identifiants

Citer

Imed Laaridh, Corinne Fredouille, Alain Ghio, Muriel Lalain, Virginie Woisard. Automatic Evaluation of Speech Intelligibility Based on i-vectors in the Context of Head and Neck Cancers. Interspeech, Sep 2018, Hyderabad, India. pp.2943-2947, ⟨10.21437/interspeech.2018-1266⟩. ⟨hal-01962170⟩
121 Consultations
203 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More