Analysis of Mispronunciation Detection and Diagnosis Based on Conventional Deep Learning Techniques

Mispronunciation Detection and Diagnosis (MDD) constitutes detecting the mispronounced chunk and diagnosing it by providing feedback about the mispronounced chunk to the model. MDD is a vital function of CAPT—Computer-Assisted Pronunciation Training, the subcategory in CALL—Computer-Assisted Language Learning. The ability of self-learning, handling huge unstructured datasets, having better evaluation metrics, and being cost-effective make deep learning models to outperform traditional models. In this work the recent deep learning methods utilizing i) input with Anti-phones, Accented content, Non-Segmentalerror, Raw input waveform, Acoustic data – Phonetic data – Linguistic data in the part of training input ii) Model suchas Transformer, Attention model, CTC-Connectionist – Temporal – Classification that are implemented on the L2-ARCTIC English dataset are analyzed. The evaluation metrics analyzed include accuracy of diagnosis and detection, F-measure. The transformer-based model shows a high F-measure value of 60.5% and a detection accuracy value of 88.2%. The hybrid CTC-ATT (CTC-Attention) model gives the second highest value of F-measure 56.08% and good performance in terms of both the diagnostic and detection accuracies of 74.96% and 75.45%, respectively. This shows that we can prefer the former method for detection and the latter method for both detection and diagnosis. Thus, based on the need for detection and diagnosis in MDD, these two models show the optimum overall performance.