Advancing Data-Constrained Spoken Language Translation

This thesis enhances spoken language translation (SLT) in data-constrained conditions by innovating in two key areas: end-to-end speech-to-text and online text-to-text translation. It introduces methods to refine speech-to-text translation by reducing training disparities between speech encoders and text decoders. It also presents scalable techniques to improve computational efficiency and augment speech data. Further, it advances online translation by employing realistic human interpretation data, proving its effectiveness over offline datasets. These contributions significantly boost both the accuracy and efficiency of SLT systems.