Speech versus keying in command and control applications

Abstract

Experimental comparisons of speech and competitor input media such as keying have, taken overall, produced equivocal results: this has usually been attributed to "task-specific variables". Thus, it seems that there are some good, and some less good, situations for utilization of speech input. One application generally thought to be a success is small-vocabulary, isolated-word recognition for command and control. In a simulated command and control task, Poock purportedly showed a very significant superiority of speech over keying in terms of higher input speeds and lower error rates. This paper argues that the apparent superiority observed results from a methodological error—specifically that the verbose commands chosen suit the requirements of speech input but make little or no concession to the requirements of keying. We describe experiments modelled on those of Poock, but designed to overcome this putative flaw and to effect a fair comparison of the input media by using terse, abbreviated commands for the keying condition at least. Results of these new experiments reveal that speech input is 10.6% slower (although this difference is not statistically significant) and 360.4% more error-prone than keying, supporting our hypothesis that the methodology of the earlier work was flawed. However, simple extrapolation of our data for terse commands to the situation where keyed commands are entered in full suggests that other differences between our work and Poock's could play a part. Overall, we conclude that a fair comparison of input media requires an experimental design that explicitly attempts to minimize the so-called transaction cycle—the number of user actions necessary to elicit a system response—for each medium.