Research on Pronunciation Error Recognition Method for English Conversation Practice Based on Artificial Intelligence Corpus

Methods relying on single-accent samples and simple rule discrimination often result in low accuracy in pronunciation error recognition. To address this challenge, a pronunciation error recognition method for English conversation practice based on an artificial intelligence corpus is proposed. Following the principles of pertinence, representativeness, and scalability, a multimodal English pronunciation corpus is constructed through preprocessing and a multi-level verbal and non-verbal annotation system. The proposed method constructs a mathematical model for the speech recognition system, performs filtering detection on pronunciation signals, and defines pronunciation feature parameters for feature extraction. The extracted features are scored against corpus patterns, normalized, and integrated through a multi-model scoring mechanism. A threshold is then used to determine the correctness of pronunciation. Experimental results show that the proposed method accurately identifies pronunciation error regions and exhibits high-position curve characteristics in Precision-Recall analysis. The F1 score increases from 0.87 to 0.99 across sample sizes from 100 to 1000. Furthermore, the method demonstrates superior classification performance in the confusion matrix, significantly improving the accuracy of pronunciation error recognition in dynamic conversational contexts.