English speech ability teaching evaluation based on active learning-based annotation and semantic expression decoupling

The evaluation of English public speaking skills has long faced challenges such as strong subjectivity, delayed feedback, and difficulty in accurately reflecting internal states like anxiety. To this end, the study developed an English speech proficiency assessment model based on active learning annotation and a semantic-expression decoupling mechanism. This model aims to enhance the objectivity and intelligence of speech anxiety identification through multimodal feature analysis, thereby providing decision-support data for teaching evaluations. The model employs active learning to prioritize annotation of high-value samples, thereby building a high-quality training dataset. Moreover, it employs a semantic-expressive decoupling architecture to separate content logic from emotional features, enabling precise identification of speech anxiety. Experiments on the UR-FUNNY and CMU-MOSEI datasets achieved an annotation accuracy of 81.88%, an anxiety recognition accuracy of 83.51%, an F1 score of 81.98%, and reduced inference latency to 67.32ms. Validation on real student datasets achieved an accuracy of 81.24%. The results demonstrate that this method exhibits significant advantages in annotation efficiency, recognition accuracy, and pedagogical interpretability, providing a viable pathway for the intelligent identification of speech anxiety and its application in teaching evaluation.