A Bayesian active-learning test for remote hearing-aid fitting

Abstract

A major problem of remote or online hearing tests is the calibration of the equipment and control of the acoustic environment: For example, an audiogram requires calibrated headphones with a wide dynamic range and a silent room. Instead, a notched-noise test alleviates these problems and gives valuable information about auditory filters that can be used to fit a hearing aid. Although absolute levels still play a role, relative levels between signal and noise are more robust as the environment does not need to be quieter than the noise. Using traditional methods, determination of auditory-filter shapes across the whole frequency range would take hours. We developed a notched-noise test that uses Gaussian Processes to maintain a probabilistic estimate of the outcome, and to select stimulus parameters based on mutual information. This active-learning method gives an estimate of auditory-filter shapes in about 30 minutes across a wide range of frequency. We demonstrate that these estimates, even when obtained with headphones for which the frequency response and exact calibration is unknown, are sufficient to give a good initial fit of a hearing aid.