Construction of Oral English Learning Model based on Multi-Head Attention with Long Short-Term Memory

One of the significant problems in the College English education of Chinese students is that traditional class teaching materials are not well supported by sufficient English learning resources, with key language points nowhere well covered systematically. The Multi-Head Attention with Long Short-Term Memory (MHA-LSTM) is proposed in this research for the construction of Oral English Learning. The MHA enable to process different segments of input sequence at similar time to extract linguistic properties. The MHA is integrated into LSTM which captures long-term dependencies including retention key phrases and grammar rules in long-sequences. The median filter is a nonlinear statistical filter which is used as preprocessing for system performance of signals and images by elimination of noise. The Mel Frequency Cepstral Coefficients (MFCC) is used as feature extraction which track dynamic characteristics of speech like positive and negative changes which are needed to recognize the prosodic features of language where the intonation and tones which are important in learning and using spoken language. The MHA-LSTM achieves 92.5% accuracy, 89.7% sensitivity, 88.3% precision and 89.1% f-measure which is better than existing algorithm.