Our laboratory specializes in the development of techniques for speech processing and natural language processing. These techniques enable machines to understand human speech, a task made increasingly possible by the rapid progress of deep learning technology. However, machine learning still requires human data annotation, which can be expensive and time-consuming. Additionally, there are numerous speech-related tasks beyond speech recognition, including speech synthesis, speech conversion, speech enhancement, and speech-based question-answering, etc. Given the vast array of such tasks, and the more than 7,000 languages spoken worldwide, it is impossible to collect large amounts of annotated data for each task and language. As a result, our laboratory has focused its recent research on reducing the amount of annotated data required for machine learning of human language. To this end, we have proposed a series of self-supervised learning (SSL) methods for speech, enabling machines to learn from unannotated data. These approaches allow machines to perform well on downstream tasks, such as speech recognition, with less labelled data, improving the efficiency and effectiveness of these processes.
