Objective prediction of the sound quality of music processed by hearing aid adaptive feedback cancellation, for normal and hearing impaired listeners

• Main Author: Manders, Alastair James
• Published: University of Southampton 2012
• Subjects: 620.23
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.580595

Adaptive feedback cancellers in hearing aids can produce unpleasant sounding nonlinear distortion artefacts (entrainment) in response to periodic or tonal inputs, including music. The aim of this study was to determine the ability of different objective metrics to predict mean quality judgments made by normal and hearing impaired listeners for music processed by hearing aid adaptive feedback cancellation. The metrics tested consisted of 'cognitive' parameters from perceptually-based audio quality models established in the literature and four new cognitive parameters described by the authors of this study. Neural networks were used to map between the values of cognitive parameters and a subjective scale of perceived quality. Network training data consisted of cognitive parameter values obtained from different excerpts of orchestral music processed by a model of a hearing aid adaptive canceller, and corresponding subjective ratings of sound quality. An optimal combination of cognitive parameters to use as network inputs was found using an extended Fourier amplitude sensitivity test (EFAST) (Saltelli et al., 1999). In an initial experiment, audio quality models were trained and validated using mean subjective ratings obtained from 26 normal hearing subjects. In a second experiment, quality model training and validation was performed using mean subjective ratings obtained from 13 subjects with cochlear hearing loss. Cognitive parameters used in the second experiment featured modifications based on subject pure-tone hearing threshold loss data to account for effects of reduced cochlear compression, frequency resolution and temporal resolution. Experimental results suggest that it possible to produce relatively accurate predictions of mean subjective quality ratings from normal and hearing impaired subjects of audio samples degraded by entrainment distortion, using perceptually-based models. This was evidenced by measured correlation scores between predicted and measured subjective ratings in excess of p=O.93. For both the normal and hearing impaired data, the most salient cognitive parameters consist of measures related to distortion loudness and the proportion of aperiodic power present in the processed signal (the latter being quantified by a new parameter, 'YINparam', which is described by the author). Cognitive parameters modelling normal hearing ears were found to be an accurate predictor of quality ratings from hearing impaired subjects; however, modifications of cognitive parameters, based on pure tone threshold data, which were used to account for supra threshold hearing loss effects did not generally improve ability to predict quality ratings from hearing impaired subjects.