Predicting Audiovisual Word Recognition in Noisy Situations: Toward Precision Audiology.

Abstract

Spoken communication is better when one can see as well as hear the talker. Although age-related deficits in speech perception were observed, Tye-Murray and colleagues found that even when age-related deficits in audiovisual (AV) speech perception were observed, AV performance could be accurately predicted from auditory-only (A-only) and visual-only (V-only) performance, and that knowing individuals' ages did not increase the accuracy of prediction. This finding contradicts conventional wisdom, according to which age-related differences in AV speech perception are due to deficits in the integration of auditory and visual information, and our primary goal was to determine whether Tye-Murray et al.'s finding with a closed-set test generalizes to situations more like those in everyday life. A second goal was to test a new predictive model that has important implications for audiological assessment. Participants (N = 109; ages 22-93 years), previously studied by Tye-Murray et al., were administered our new, open-set Lex-List test to assess their auditory, visual, and audiovisual perception of individual words. All testing was conducted in six-talker babble (three males and three females) presented at approximately 62 dB SPL. The level of the audio for the Lex-List items, when presented, was approximately 59 dB SPL because pilot testing suggested that this signal-to-noise ratio would avoid ceiling performance under the AV condition. Multiple linear regression analyses revealed that A-only and V-only performance accounted for 87.9% of the variance in AV speech perception, and that the contribution of age failed to reach significance. Our new parabolic model accounted for even more (92.8%) of the variance in AV performance, and again, the contribution of age was not significant. Bayesian analyses revealed that for both linear and parabolic models, the present data were almost 10 times as likely to occur with a reduced model (without age) than with a full model (with age as a predictor). Furthermore, comparison of the two reduced models revealed that the data were more than 100 times as likely to occur with the parabolic model than with the linear regression model. The present results strongly support Tye-Murray et al.'s hypothesis that AV performance can be accurately predicted from unimodal performance and that knowing individuals' ages does not increase the accuracy of that prediction. Our results represent an important initial step in extending Tye-Murray et al.'s findings to situations more like those encountered in everyday communication. The accuracy with which speech perception was predicted in this study foreshadows a form of precision audiology in which determining individual strengths and weaknesses in unimodal and multimodal speech perception facilitates identification of targets for rehabilitative efforts aimed at recovering and maintaining speech perception abilities critical to the quality of an older adult's life.