Applying Machine Learning to Speech Processing in Hearing Aids

Abstract

<p>The neurophysiological basis of sensorineural hearing loss is thought to be hair cell damage or stria vascularis atrophy in the cochlea. The normal cochlea is responsible for a very complex, dynamic, nonlinear analysis and coding of acoustic signals, which is distorted by cochlear impairment. To overcome hearing loss, a typical hearing aid provides linear gain or some simple form of dynamic compression. However, such simple processing cannot fully compensate for the effects of cochlear impairment. In this thesis, machine learning is used to investigate more optimal speech processing schemes for hearing aids.</p> <p>A model of the auditory periphery is utilized to develop a set of neural predictors of human speech intelligibility. These are shown to have similar accuracy to acoustic predictors of intelligibility such as the articulation index. The neural predictors are then used as error metrics in a machine learning framework to train simple linear and compressive hearing aid algorithms. The results are consistent with empirically derived prescriptions for hearing linear gain and compression.</p> <p>It thus appears that to develop speech processing algorithms that provide greater benefits than those currently available in hearing aids, it is necessary understand more fully the distortions that are occurring in the cochlea due to hearing loss and to develop processing algorithms that specifically target compensation of these distortions. An analysis of the differences in compression, suppression and adaptation in the normal and impaired cochlea is performed using the model of the auditory periphery, and specific distortions are quantified. From this analysis, several speech processing algorithms are proposed that may more fully compensate for the effects of cochlear impairment on the neural representation of speech.</p>