Nativeness Classification with Suprasegmental Features on the Accent Group Level

摘要

We present a novel approach to discriminating native and nonnative utterances based on suprasegmental features extracted at the Accent Group (AG) level. Past studies have shown modeling a set of shared intonation patterns across AGs to be effective in predicting local fO contour shapes. Here we demonstrate that AG level prosodic features are also effective in nativeness classification. The proposed suprasegmental feature set is very low dimensional, and is derived from fO and energy contours across the AG, as well as normalized duration of the syllables within each AG. A Random Forest back end classifier is used to combine AG level scores from GMM and Decision Tree models, producing nativeness scores at the utterance level. The proposed prosodic nativeness classifier achieves 83.3% accuracy for 2-AG utterances and 89.1% accuracy for 3-AG utterances, exceeding a baseline Gaussian Supervector system's performance by more than 10% absolute. The vastly lower dimensionality of the proposed feature set relative to the baseline method suggests the importance of suprasegmental features over traditional spectral cues in contributing to the perceived nativeness of a learner's language.