Description:

Linguistics Colloquium Series: Tessa Bent

Affiliation: Indiana University

Topic: Relating pronunciation distance metrics to children and adults’ accent perception

Substantial variability in the speech signal arises from within- and across-talker factors (e.g., speaking style, health status, gender, dialect, native language). Rather than viewing this variability as noise, there is now ample evidence that listeners encode both linguistic and socio-indexical information in highly detailed cognitive representations and that these information sources interact in ways that typically give rise to robust speech comprehension. However, some sources of variability can cause substantial decrements in speech comprehension, particularly for listeners with developing linguistic systems. For example, early school-aged children show lower word identification accuracy under both environmental- and source-related listening challenges (e.g., speech in noise or unfamiliar accents) compared to adults. In fact, fully mature word identification abilities with non-local accents may not emerge until adolescence. One limitation of this research is that it typically does not include quantifications of how the non-local accents differ from local accent norms. Thus, we are evaluating how quantitative pronunciation distance metrics for native and nonnative varieties of English predict word recognition accuracy and perceived accent distance from the local standard. For word recognition in optimal listening conditions, productions that diverged more substantially from the ambient accent based on a segmental distance metric (i.e., Levenshtein distances) were harder for children and adults to identify with similar effects of distance for the two age groups. In noise-added conditions, children showed greater decrements in word recognition accuracy with increasing distance from the local accent than adults, suggesting that these more challenging mappings may have exceeded the cognitive resources (e.g., working memory) available to children. For perceived accent distance as measured by an accent ranking task, a weighted Levenshtein measure was the best predictor, but unweighted Levenshtein and dynamic time warping measures also significantly contributed. Our accent ranking task also suggested that even 6-year-old children are sensitive to both non-local native and nonnative accents, a result that contrasts with earlier claims that early school-aged children show limited abilities to distinguish among different regional accents. In ongoing work, we are incorporating suprasegmental measures, sentences with more varied syntactic structures and lengths, and a wider range accents. Establishing reliable and predictive measures of pronunciation distance is key for understanding developmental changes in speech processing with less familiar accents. [Research supported by the National Science Foundation - Award Numbers: 1941691, 1941662, and 1461039].

Recording of the talk will be available upon request. Please contact Linguistics@northwestern.edu