Review of Recent Systems for Detecting and Diagnosing Pronunciation Errors

Nowadays, Computer Aided Language Learning (CALL) frameworks have attracted a lot of attention because of their capability, adaptability, and flexibility in improving people's language skills. The field of Mispronunciation Detection and Diagnosis (MDD) is considered one of the applications of these frameworks and has recently benefited from the rapid innovation spawned in deep learning models and different acoustic, phonetic, and linguistic features. The various manner of combinations of different features and deep learning architectures for different MDD systems require huge amount of annotated textual data to be pretrained and hence achieved good performances. However, providing a large amount of annotated textual data for pre-training is a tedious and time-consuming operation. For this, a pre-trained acoustic model from unannotated data can be a good alternative. In this article, we present different MDD systems categorized according to the pre-training mode: pre-trained approaches without annotated text data and pre-trained approaches with annotated text data. Experimental results, advantages, disadvantages and additional improvements for each method explored for each category are presented and discussed.