Unsupervised Discovery of Non-Categorical L2 Error Patterns Using Wav2Vec2.0 Code Vectors

L2 pronunciation is shaped by the interaction of two sound systems, which makes their identity more complex than a single phoneme category. The non-categorical nature demands assessment at a level finer than phonemes. As the granular requirement is highly labor-intensive, unsupervised methods emerged. Nevertheless, they either reverted to categorical diagnosis or used the supervised and phoneme-prescribed feature phonetic posterior-gram (PPG). Alternatively, this study adopts the unprescribed and unsupervised feature, the Wav2Vec2.0 code vector, to locate sub-phonemic variations. We first verify the features’ L2 discernability by comparing their frequency across single-speaker data of L1 (CMU ARCTIC) and L2 (L2 ARCTIC). Clustering is performed on frequency vectors to test their separability on account of nativeness. Subsequently, sub-segmental patterns are analyzed among segmentally identical error samples in L2 Korean English NIA 037 data. After cataloging segmental errors detected by the model finetuned with L1 TIMIT, their corresponding code vector sequences are extracted by referencing the forced alignment result. We then derived dominant patterns of the sequences and compared them against L1 reference materials constructed from TIMIT. Phoneme-code vector co-occurrence probability and code vector clustering were each used to check their attributes and uniqueness. The result confirmed the discernability, followed by linguistically interpretable common traits across patterns. (1) They formed a gradient error continuum along the changed articulatory value, reflecting the non-categorical nuanced understanding. (2) This trait is highlighted by intermediary typology assuming opposite values in two codebooks which was also rare in L1 for being L2 specific. Lastly, (3) distribution skewed towards the most approximate sound in the learner’s L1, from which the patterns’ complexity stems.