Improving Automatic Speech Recognition for Low-Resource Language by Data Augmentation

Automatic speech recognition (ASR) is one of the emergency tasks in human-computer interaction. There are many studies work in the field of building network architecture to deal with this task. While data augmentation was deeply discovered in computer vision, it is a big lag behind in the field of speech. Large data collection is not trivial, and in some cases it is impossible. The problem with data size is even more serious in some low-resource languages, such as Vietnamese. This study focuses on the data augmentation approach to deal with the small-size datasets to help the deep learning network better coverage in the ASR task. The experiment results on various configures of the VIVOS dataset, and two variations of the Conformer network architecture show that our proposed method gets promising improvement.