Deep learning-based intelligent diagnosis and adaptive training system for university english oral proficiency

Abstract

Abstract The development of oral English proficiency among university students remains constrained by limitations in traditional assessment approaches and the scarcity of personalized training resources. This study proposes an intelligent diagnosis and adaptive training system integrating deep learning technologies with multidimensional oral proficiency assessment theory. The diagnostic component employs a multi-task learning framework combining CNN-LSTM architectures with attention-based multimodal fusion to simultaneously evaluate pronunciation accuracy, fluency, vocabulary-grammar complexity, and content coherence from continuous speech samples. The adaptive training component utilizes reinforcement learning to optimize personalized practice recommendation sequences based on dynamically tracked learner competence states. Experimental validation on a corpus of 4,374 recordings from 486 university students, using a strict speaker-independent data partitioning protocol, demonstrates that the diagnostic model achieves correlation coefficients of 0.887, 0.862, 0.824, and 0.793 with human expert ratings across the four assessment dimensions respectively. Controlled training experiments over eight weeks reveal that learners receiving reinforcement learning-optimized content sequencing attain normalized learning gains approximately 2.2 times higher than those following fixed curricula (Cohen’s d = 1.34, 95% CI [0.87, 1.81]), while maintaining stronger engagement throughout the training period. These findings provide preliminary evidence supporting the potential of the proposed framework for scalable, individualized oral English instruction in higher education contexts, though further validation across broader populations and longer time periods is warranted.