An adaptive sampling procedure for speech perception experiments

Abstract

Synthetic speech perception experiments may make use of several acoustic dimensions in order to adequately model listeners' perception; however, the number of stimuli increases exponentially as dimensions are added. A relatively large number of identification responses per stimulus are needed in the vicinity of category boundaries in order to model the boundaries with reasonable accuracy. Fewer responses per stimulus are needed to model portions of the stimulus space where a single response category predominates. Rather than collecting the same number of responses for each stimulus, an experiment can therefore be shortened via adaptive sampling. An adaptive sampling procedure is described. After an initial pass through the stimuli, the procedure uses a logistic regression model to select stimuli to resample in subsequent rounds. Results of simulations indicated that the number of trials in the experiment could be reduced by a third without substantially affecting the results.