A grading evaluation method for English oral pronunciation errors based on deep neural networks

In this paper, a deep neural network-based grading method for English oral pronunciation errors is proposed. First, English oral pronunciation signals are pre-processed and MFCC feature vectors are extracted. Second, Hidden Markov model is used to construct an acoustic model, with deep neural network used to predict the state probability distribution of acoustic feature vectors, replacing the observation probability of the acoustic model. Then, a language model is constructed to obtain the probability of word order, and combined with an acoustic model to build a search network. Viterbi algorithm was used to decode and find the phoneme state sequence. Finally, based on the reference phoneme sequence, the degree of pronunciation errors are calculated, compared with a threshold, and achieved graded evaluation. The results indicate that the AUC value of the proposed method is close to 1, and the F1 value is above 0.95, indicating a high accuracy of the evaluation.