Speech, Hearing and Language: work in progress

One of the main problems in carrying out psychoacoustic experiments is the time required to measure a single threshold. In this study we compare the accuracy of threshold estimation in a 2I2AFC task for detecting a 2kHz tone in either a broadband noise or a notched-noise. Tone thresholds were estimated in three normal-hearing listeners using either a Levitt procedure to track 79% correct, or a maximumlikelihood estimation (MLE) procedure to track 70, 80 or 90% correct. Given the chosen parameters for the different procedures, the MLE procedure proved to be approximately 2.5 times faster at estimating masked thresholds than the Levitt procedure. Only thresholds using the 70% MLE procedure were significantly different in magnitude from those obtained using the Levitt procedure. To test the repeatability of the measurements the standard deviations (SD) of the threshold were calculated. Statistical analyses show smallest SDs for the Levitt and 90% MLE procedures, with significantly larger SDs for the 70% and 80% MLE. Introduction Of major concern in designing psychoacoustic experiments is not just the issue to be investigated, but also the time available for the experimental tests to be carried out. The desire for efficient threshold estimation has led to the adoption of several different procedures typically using one, two or three alternative forced choice techniques. Probably the most widely used procedure in psychoacoustics is the adaptive technique based on the transformed up-down procedure described by Levitt (1971). In this technique the initial stimulus is set “above” threshold and subsequent presentation levels are governed by (a) the step-size used and (b) the response to the current stimulus. Correct responses make the task harder by the given step-size and incorrect responses make the task easier. The choice of step-size (fixed or variable), and the patterns of correct/incorrect responses leading to a reversal are described in detail by Levitt (1971). More recently, considerable interest has been shown in other threshold estimation techniques. In particular, with the advent of increased computing power in laboratories and clinics, the maximum-likelihood estimation (MLE) procedure (Hall, 1968) has become more widely used. The basic premise behind this procedure is that the experimenter assumes a parametric form for the psychometric function, and after each new response the values of the parameters are computed that “...maximise the probability of the set of responses that have been obtained, given the set of stimuli that have been presented”. Several studies have compared maximum-likelihood threshold estimation techniques with other techniques in either computer simulations, or empirical measurements (e.g. 1 Centre for Human Communication & Deafness, Faculty of Education, University of Manchester, Oxford Road, Manchester M13 9PL