| Three factors have been identified as contributing to the speechreading of sentences. These are the skill of the observer, the visual intelligibility of the talker, and the sentence itself. This study was designed to investigate the role of the sentence in the speechreading paradigm, and to explore possible characteristics of sentences that contribute to their intelligibility, independent of the talker and the observer. Four studies were conducted to examine the sentence effect. The first study was a replication of Demorest & Bernstein's generalizability study (1992), which identified the sentence effect as the most dominant of the three factors. The generalizability study revealed that 13.5% of the variance was accounted for by the observer, 0.0% of the variance was accounted for by the talker, and 33.2% was accounted for by the sentence. The replication, involving new talkers and a different scoring method, agreed well with Demorest & Bernstein (1992). There was a talker by sentence interaction that accounted for 2.5% of the variance, and error accounted for the remaining 50.8%. Study II examined the durability of the sentence effect across talkers, with the observer effect removed. In addition to the two new talkers used in this research, data from Hinkle (1978) and Demorest & Bernstein (1992) were included, yielding a total of five talkers. Results indicated a strong sentence effect, accounting for over 50% of the variance. The focus of Study III was the effect of observer performance level on the sentence effect. Participants were divided into high, medium and low performance groups, and correlations were calculated within each talker across sentences. Again, the sentence effect accounted for up to half of the variance, consistent with the strength of the sentence effect seen in Studies I and II. Study IV sought to discover characteristics of sentences which could predict their percent keyword correct scores. Factors involving the likelihood of the sentence, the frequency of the initial keywords in the sentence, the number of keywords per sentence, and the ratio of function to content words were significant predictors of speechreading performance, resulting in R-square values approaching 0.5. |
