On Knowledge Distillation for Direct Speech Translation
This research addresses the challenge of improving translation quality in direct speech translation for users who need real-time language conversion, representing an incremental improvement in knowledge transfer techniques.
This paper explores knowledge distillation techniques for direct speech translation (ST), a complex task benefiting from knowledge transfer from ASR and MT. It compares different distillation solutions for sequence-to-sequence ST tasks and analyzes methods to mitigate potential drawbacks while preserving translation quality benefits.
Direct speech translation (ST) has shown to be a complex task requiring knowledge transfer from its sub-tasks: automatic speech recognition (ASR) and machine translation (MT). For MT, one of the most promising techniques to transfer knowledge is knowledge distillation. In this paper, we compare the different solutions to distill knowledge in a sequence-to-sequence task like ST. Moreover, we analyze eventual drawbacks of this approach and how to alleviate them maintaining the benefits in terms of translation quality.