A Flexible Question-and-Answer Task for Measuring Speech Understanding

Abstract

This report introduces a new speech task based on simple questions and answers. The task differs from a traditional sentence recall task in that it involves an element of comprehension and can be implemented in an ongoing fashion. It also contains two target items (the question and the answer) that may be associated with different voices and locations to create dynamic listening scenarios. A set of 227 questions was created, covering six broad categories (days of the week, months of the year, numbers, colors, opposites, and sizes). All questions and their one-word answers were spoken by 11 female and 11 male talkers. In this study, listeners were presented with question-answer pairs and asked to indicate whether the answer was true or false. Responses were given as simple button or key presses, which are quick to make and easy to score. Two preliminary experiments are presented that illustrate different ways of implementing the basic task. In the first experiment, question-answer pairs were presented in speech-shaped noise, and performance was compared across subjects, question categories, and time, to examine the different sources of variability. In the second experiment, sequences of question-answer pairs were presented amidst competing conversations in an ongoing, spatially dynamic listening scenario. Overall, the question-and-answer task appears to be feasible and could be implemented flexibly in a number of different ways.

Keywords: dynamic conversations; realistic tests; speech comprehension.

Figure 1.

(a) Psychometric functions for each listener (gray lines) and the mean psychometric function (black lines and circles). (b) Mean psychometric functions for each category type. (c) Mean psychometric functions based on the first half (circles) and second half (squares) of the trials completed by each subject at each SNR. (d) Mean psychometric function in units of d′ (circles) and bias (squares). Error bars, where shown, represent across-subject standard deviations.

Figure 2.

Spatial and temporal configuration of the targets and maskers in the dynamic listening condition. The nine horizontal lines represent the nine stimulus positions (from −60° to +60° azimuth), of which three were potential target locations (−30°, 0°, +30°), and the remaining six were occupied by three pairs of masker talkers engaged in conversation. The shaded bars indicate the times during which a particular talker was speaking (colors: targets; gray: maskers). In this example, three questions and answers (labeled Q1, A1, etc.) out of the sequence of 12 are shown. The fixed condition was identical except that all questions and answers came from one of the three target locations.

Figure 3.

(a) Psychometric functions for the fixed and dynamic listening conditions (averaged across all locations and all listeners). (b) Psychometric functions in units of d′ (circles) and bias (squares) for the fixed and dynamic listening conditions (averaged across all locations and all listeners). Error bars in both panels show across-subject standard deviations.