Automated Assessment of L2 English Oral Proficiency Using Transformer Models

The automated assessment of oral proficiency in second language acquisition holds significant importance for enhancing learning efficiency and represents a key research direction in the intelligent transformation of language teaching. However, current methods still exhibit notable limitations in terms of accuracy, generalizability, and practicality. Approaches relying on manually designed features are costly, exhibit weak generalization, and struggle to capture complex grammatical and semantic information. Furthermore, methods based on speech recognition or text-only models are often constrained by transcription quality and computational resources, showing limited adaptability to non-standard pronunciation, accents, and disfluent speech, while also failing to utilize acoustic features effectively. In this paper, we propose an automatic oral assessment method based on the Transformer model Wav2vec 2.0 XLS-R which utilizes a pre-trained model for efficient feature extraction to evaluate the oral performance of second language learners. The proposed approach can effectively improve the reliability and validity of diagnosing spoken errors among language learners.