A study on fine-tuning wav2vec2.0 Model for the task of Mispronunciation Detection and Diagnosis

Mispronunciation detection and diagnosis (MDD) technology is a key component of computer-assisted pronunciation training system (CAPT). The mainstream method is based on deep neural network automatic speech recognition. Unfortunately, the technique requires massive human-annotated speech recordings for training. Due to the huge variations in mother tongue, age, and proficiency level among second language learners, it is difficult to gather a large amount of matching data for acoustic model training, which greatly limits the model performance. In this paper, we explore the use of Self-Supervised Pretraining (SSP) model wav2vec2.0 for MDD tasks. SSP utilizes a large unlabelled dataset to learn general representation and can be applied in downstream tasks. We conduct experiments using two publicly available datasets (TIMIT, L2-arctic) and our best system achieves 60.44% f1-score. Moreover, our method is able to achieve 55.52% f1-score with 3 times less data, which demonstrates the effectiveness of SSP on MDD1.