Estimation of auditory filter shapes across frequencies using machine learning

Abstract

When fitting a hearing aid, the level-dependent gain prescribed at each frequency is usually based on the hearing loss at that frequency. This often results in reasonable fittings for a typical cochlear hearing loss, but may fail when the individual frequency selectivity and/or loudness growth are different from what would be typical for that hearing loss. Individualised fitting based on measures of frequency selectivity might be useful in improving a fitting, for example by reducing across-channel masking. A popular measure of frequency selectivity is the notched-noise method, but this test is time-consuming. To reduce testing time, Shen and Richards (2013) proposed an efficient machine-learning test that determines the slope of the skirts of the auditory filter (p), its minimum response for wide notches (r), and detection efficiency (K). However, their test did not determine asymmetries in the auditory filter, which are important to consider during fitting to reduce across-channel masking. The test proposed here provides a time-efficient way of estimating the auditory filter shape and asymmetry as a function of center frequency. The noise level required for threshold is estimated for a tone with frequency fs presented at 15 dB SL in nine symmetric or asymmetric notched noises with notch edge frequencies between 0.6 and 1.4 fs. Using only narrow to medium notch widths provides good information about the tip of the auditory filter, which is of most importance in determining across-channel masking for speech-like signals (but the tail is not well defined). The nine thresholds for a given fs can be used to fit an auditory filter model with three parameters: the slopes of the lower and upper sides (pl, pu) and K. In practice, these model parameters are estimated as a continuous function of fs, and fs is varied across trials over the range 0.5-4 kHz. The stimulus parameters on a given trial (fs, notch condition, noise level) are chosen to maximally reduce the uncertainty in the model parameters, exploiting the covariance between thresholds for adjacent values of fs. Six subjects have been tested so far. The whole procedure took about 45 minutes per ear. The lower slopes typically corresponded with values expected from the audiogram and a cochlear hearing loss. The upper slopes were steeper in some cases, although not necessarily across the whole frequency range.

Reference Shen, Y., and Richards, V. M. (2013). "Bayesian adaptive estimation of the auditory filter," J. Acoust. Soc. Am. 134, 1134-1145.