A Bayesian active-learning approach for obtaining notched-noise data
* Presenting author
A popular method for characterising frequency selectivity is the notched-noise method, but this is time consuming, requiring measurement of masked thresholds in a two-dimensional space with the lower and upper spectral edges of the notch (Fl and Fu) as parameters. We used a Gaussian Process (GP) for classification, with the masker level as a third dimension, to determine the probability of a fixed pure-tone signal being detected. The GP used the assumptions that: the detection probability is correlated for similar notch parameters; wider notches increase the detection probability; decreasing the noise level increases the detection probability according to a cumulative Gaussian psychometric function. The masker parameters chosen for the next trial were chosen to yield the highest mutual information about the parameter space. The thresholds predicted by the GP after 100 trials using normal-hearing listeners changed more rapidly with changes in Fu than with changes in Fl, as expected from the Zwicker-Fastl excitation-pattern model. The predictions of the GP were used to fit rounded-exponential auditory filters. The slope parameters determined in this way agreed with published values. The method may provide a quick test for determining asymmetries in the auditory filter shapes of hearing-impaired listeners.