1. This is a very important point, in that all of the studies of the effects of rate manipulations (including all of the studies from our research group as well as the Apple, Streeter and Krauss, 1979 study reviewed in this paper) are crude in this regard. That is, whereas it is very possible to use the computer synthesis scheme to produce voices that are increased or decreased in rate the same way natural variations occur, with a differential amount of change in vowels and consonants, or at clause boundaries or phrase boundaries, etc., this would require gathering considerable data first to determine the properties of such natural changes in rate and it also requires the incorporating of such data into the synthesis process. Although we are now beginning some work in that direction, all of the studies of rate manipulations up to now could just as well have used the mechanical rate-changer device, in that the computer synthesis method that was employed changed all segments of the utterance uniformly.
2. This was determined by taking one minus the Wilks' lambda value (1-lamda) from the multivariate analysis of variance. See Cooley and Lohnes (1971, pp 223–242) for the rationale of this percent of variance statistic.
3. Speakers were randomly assigned to treatment combinations and an analysis of variance was computed on pre-manipulation pitches and rates, and it was found that a preponderance of slow speakers were, unfortunately, assigned to the slow rate condition. Apparently the study had already been completed when this check was run, since, rather than altering the speaker assignments for a more careful experimental control, analysis of covariance was used to statistically accommodate the problem. Covariance, of course, assumes a linear relation. Smith's finding of a curvilinear relation between benevolence adjectives and speaking rate suggests that the covariance analysis may not have solved all of the problems.
4. The experimenter worked with each speaker in repeated tries to reach the maximum and minimum rates that still sounded like real speech.
5. These data were gathered and analysed by Brett E. Murri as a project for an experimental psychology class, and the data for the other ratings of these voices and the gathering and synthesis of the voices are from a similar project by Mark A. Stewart and Stephen Stewart. A more detailed analysis is forthcoming in a jointly authored paper.