Suchergebnisse

Both analytic and Monte Carlo methods were used to show that empirical estimates of the standard error of an averaged correlation agree quite well with theoretical expressions derived from Pearson's asymptotic equations. Where discrepancies did exist, particularly with small sample sizes, the theoretical formula slightly overestimated the actual (empirical) standard error, thus making a test of significance conservative. It was shown that by testing repeatedly with a small sample and averaging the resulting relations, one may obtain precision equivalent to single estimates from much larger samples. Furthermore, the greatest increase in precision occurred when the population correlation was small, the situation where greatest precision is needed. These results should prove of value for human factors evaluations, which often have limited subject resources.

This research investigated the differential impact of synthetic voice quality and text difficulty on comprehension of extended prose. Sixty participants listened to five easy and five difficult passages in one of three speech modes: natural speech, VOTRAX (low intelligibility), or DECtalk (high intelligibility). Comprehension of DECtalk was equal to that of natural speech, whereas comprehension of VOTRAX was significantly poorer than with natural speech or DECtalk. Subjects were also asked to shadow passages of each speech type as a measure of resource processing demands. It was found that shadowing accuracy was significantly better for natural speech than for DECtalk and shadowing of DECtalk was markedly superior to that of VOTRAX. The results of this study suggest that resource-demand measures alone may not be appropriate to predict performance in practical applications. Specifically, overall comprehension may not suffer despite on-line losses in processing. These findings also point to a differential allocation of cognitive resources by speech synthesizers of differing intelligibility.