14th Speech in Noise Workshop, 12-13 January 2023, Split, Croatia 14th Speech in Noise Workshop, 12-13 January 2023, Split, Croatia

P02Session 2 (Friday 13 January 2023, 09:00-11:00)
Evaluating the intelligibility of hearing-aids in realistic scenes: The Hearing-Aid Listening Test

Gaston Hilkhuysen, Tim Green, Stuart Rosen, Mark Huckvale
Department of Speech Hearing and Phonetic Science, University College London, UK

Evaluating the intelligibility one obtains with commercial hearing aids (HAs) is complex. Each HA employs a combination of digital signal processing strategies with implementations differing across manufactures and models. Measurements of the aids’ physical properties can be helpful to quantify performance but are difficult to relate to realistic listening scenes. Various intelligibility metrics claim they account for the processing found in HAs but have not been validated under realistic conditions. Consumers, on the other hand, demand the HA that is best fit to their individual needs. This raises the following questions. 1) Given a hearing-impaired listener and a listening scene, which HA provides optimal intelligibility? 2) Can intelligibility metrics be used for HA selection?

To answer these questions, we recorded the outputs of six pairs of commercial HAs fitted to a manikin placed in four locations: an anechoic room, a kitchen, a car and a restaurant. IEEE test sentences were played at approximately 1 m distance from two directions, either facing or to the left of the manikin. Speech-shaped noise was played from the facing loudspeaker in the anechoic chamber, while elsewhere background noises typical to the location were presented.

Each HA was fitted to the same mild-to-moderate sloping audiogram. As two baseline conditions we recorded the sentences in an unaided condition and additionally processed these unaided signals with NAL-R1 amplification. This led to a database with 64 test conditions: 8 processors, 4 scenes and 2 target speaker positions.

We subsequently measured the intelligibility performance of 16 HA users with losses similar to the target audiogram in the 64 conditions. Detailed results will be presented on the poster, but highlights include the facts that no single HA was best for all listeners or for all scenes; and frequently NAL-R1 amplification outperformed all the commercial HAs. Furthermore, variation in intelligibility performance across aids was not well explained by a signal-based metric designed for prediction of intelligibility by hearing-impaired listeners.

Last modified 2023-01-06 23:41:06