L1-Aware Phoneme-Level Pronunciation Assessment of Uzbek L1– English L2 Speech Using wav2vec 2.0 Acoustic-Phonetic Representations

Automatic pronunciation assessment is especially important for Uzbek learners of English because differences between the Uzbek and English phonological systems can produce systematic L1-induced deviations in segmental and prosodic realization. This paper proposes an interpretable pronunciationassessment model in which wav2vec 2.0 is used as a self-supervised acousticphonetic representation extractor, while the assessment stage is formulated as a phoneme-level scoring problem rather than a conventional speech-recognition problem. For each target phoneme, the model combines four complementary indicators: posterior evidence for the canonical phoneme, distance between learner and reference embeddings, duration/prosodic conformity, and an Uzbek-L1-specific confusion penalty. These components are fused by a bounded logistic scoring function and then aggregated hierarchically from phoneme to word and utterance levels. The mathematical formulation is designed to preserve interpretability: every final score can be decomposed into acoustic, temporal, prosodic, and L1interference contributions. The proposed formulation is suitable for low-resource Uzbek L1–English L2 corpora because most representation learning is transferred from wav2vec 2.0, while only the calibration and scoring layers require taskspecific labeled data. The model provides a technical basis for future computerassisted pronunciation training systems that can produce both a numerical score and error-oriented feedback.