Comparing Levenshtein distance and dynamic time warping in predicting listeners’ judgements of accent distance

Abstract

Non-native accent ratings are related to segmental and holistic acoustic deviations from listeners’ home accents (Bartelds et al., 2020). This study builds on prior work by including both non-ambient native and nonnative accents, using a different perceptual task in which listeners ranked talkers based on their perceived distance from Standard American English, and employing analyses that account for collinearity between distance metrics. Listeners (n = 52) completed two rankings where all talkers produced the same sentence and one where each talker produced a unique sentence. Phonemic and holistic acoustic distances between the nine non-ambient accents and Midland American English were quantified using Levenshtein distances and dynamic time warping (DTW), respectively. Results from separate linear mixed effects models showed that both DTW and Levenshtein distances contributed to perceptual distance. Sentence and its interaction with DTW/Levenshtein distance were not significant. The model predicting perceptual rankings from DTW was a better fitting model than the model with Levenshtein distances. Because DTW captures both segmental and suprasegmental distance, it may better predict listeners’ similarity judgements among English varieties than a metric only capturing phonemic distance. [Funded by the National Science Foundation (1941691; 1941662) and The Ohio State University Center for Cognitive and Brain Sciences.]