A validation platform for binaural speech intelligibility metrics

Abstract : Numerous studies have addressed speech intelligibility in situations where the target speaker and interfering sources are at different angular positions relative to the listener. Effects of number, location and type of interferers, the influence of room acoustics as well as the listeners hearing status or directional signal processing are well documented. Authors have additionally proposed a variety of speech intelligibility metrics typically developed based on a subset of available data. Although data is indispensable during metric development, there is a risk that metrics fail to generalize to other data. Within the E-Lobes project, we are developing a test bench to validate a variety of existing speech intelligibility metrics on existing data, unused during their development. The test bench consists of six interconnected modules: 1) audio sources; 2) listening conditions; 3) listener properties; 4) intelligibility scores; 5) intelligibility metrics and 6) evaluation statistics. Varying the elements in each module allows one to inspect the reliability and validity of metrics. It also identifies conditions where metrics’ predictions diverge, hence suggesting conditions for subsequent intelligibility testing. Some preliminary results will be presented.
Type de document :
Poster
3èmes Journées Perception Sonore, Jun 2017, Brest, France. 2017, 〈http://www.univ-brest.fr/JPS2017〉
Liste complète des métadonnées

http://hal.univ-brest.fr/hal-01559381
Contributeur : Vincent Koehl <>
Soumis le : lundi 10 juillet 2017 - 16:20:33
Dernière modification le : mercredi 12 juillet 2017 - 10:35:51

Identifiants

  • HAL Id : hal-01559381, version 1

Collections

Citation

Gaston Hilkhuysen, Tim Green, Stuart Rosen, Mark Huckvale. A validation platform for binaural speech intelligibility metrics. 3èmes Journées Perception Sonore, Jun 2017, Brest, France. 2017, 〈http://www.univ-brest.fr/JPS2017〉. 〈hal-01559381〉

Partager

Métriques

Consultations de la notice

32