Uncovering between-categorical details in L2 pronunciation errors using Wav2Vec2.0 code vectors*
Eunsoo Hong, Sunhee Kim, Minhwa Chung · Phonetics and Speech Sciences · 2024
L2 pronunciation errors, influenced by L1 phonetic transfer, often span the boundaries of two or more canonical phonemes. This phenomenon necessitates moving beyond the categorical judgments by analyzing pronunciation using units more granular than phonemes. While previous studies proposed phonemic posterior-grams (PPG) in this context, these features remain tied to a categorical framework when derived through phonemic association. This study employs Wav2Vec2.0 code vectors as an alternative description unit, granted that the self-supervised learning (SSL) representation is free from external phonemic stipulation. We first evaluate whether code vectors encode L1 and L2 speech differently and use them to identify patterns among categorically identical error samples. To this end, the used inventory in single-speaker data of L1 (CMU ARCTIC) and L2 (L2 ARCTIC) is first compared. Then, substitution errors are selected from our L2 analysis NIA037 data using the model finetuned with TIMIT. For each error type, sub-categorical inspection concerns deriving dominant patterns from corresponding featural sequences and interpreting them using raw code vector clustering results and phoneme-code vector co-occurrence probabilities, both constructed from the L1 reference data TIMIT. The findings revealed that, beyond discerning L2 speech, code vectors can effectively quantify the between-categorical positions of error patterns in the pronunciation spectrum. Specifically, Euclidean distances between pattern vectors reflected phonetically grounded variation phenomena documented in existing L2 literature.