A signal detection theory approach to predicting auditory detection performance.

摘要

Accurate prediction of human auditory detection performance on psychophysical tasks would contribute to improved acoustic signal coding strategies for many applications, such as hearing aids, perceptual coding schemes, and improved human-system interfaces. Historically, signal detection theory has been applied directly to stimuli or to the output of basic auditory models. Unfortunately, theoretical predictions and experimental measures of detection performance rarely agreed, with attempts to reconcile the discrepancies often neither physiologically-based nor universally applicable.;This thesis addresses the problem of inaccurate theoretical predictions by applying a signal detection theory analysis to the outputs of several computational auditory models. Proof of concept was provided when a simple simultaneous masking task was analyzed using a functional auditory model. The theoretical results, though not accurate predictions of human performance, nonetheless were generally more accurate than traditional approaches. Rather than eliminate the remaining discrepancies by adding a Gaussian noise source, as was common previously, the differences were partially explained in physiological terms.;The effects of stimulus uncertainty, stimulus level, and neural uncertainty, as well as the specific implementation of each auditory model, on detection performance were evaluated. No single effect completely accounted for the discrepancies between theoretical and experimental results. However, when neural uncertainty, representing "internal noise", and stimulus phase uncertainty were incorporated into the problem, remarkably accurate predictions of detection performance were obtained. The success of this integrated approach in predicting simultaneous masking data, as well as in accurately predicting the independence between signal level and signal detectability, is promising and indicates that this technique will be viable for a wide range of applications.