This paper addresses the need for English speaking training in a digital education environment by designing an intelligent English speaking training system based on deep learning. The system employs a semantic understanding model that integrates role information and historical dialogue context, utilizing a BERT-BiLSTM-CRF joint framework to achieve intent recognition and slot value filling. In the speech preprocessing stage, the system innovatively applies spectral entropy-based endpoint detection (VAD) to optimize the processing of low-energy speech signals, and combines pre-emphasis and Hamming window framing techniques to enhance recognition robustness. On the ATIS dataset, the system achieves an intent recognition accuracy of 99.19% and a slot filling accuracy of 97.24%, representing improvements of 0.7% and 1.1% over the best baseline, respectively. System performance testing shows that in real teaching interactions, the average response latency is 1024.2 ms, with 98% speech recognition accuracy, 98% task completion rate, and 92% pronunciation correction rate. In educational empirical studies, students’ oral English scores significantly improved from 71.06 ± 15.99 points to 88.31 ± 8.54 points (+24.25%), the failure rate decreased from 24.51% to 0%, and the excellent rate (>90 points) increased from 16.67% to 48.04%. The learning attitude questionnaire showed that the number of students who fully agreed with “fluent English speaking” increased from 41 to 80 (+95.1%), and the willingness to persist in daily training increased by 252.4% (from 21 to 74 students). The study indicates that the system effectively enhances oral training efficiency through deep semantic understanding and multimodal interaction design, providing technical support for digital English teaching.