Enhancing End-to-end Multilingual Medical Speech Translation via Terminology Injection Mechanism

Speech-to-text translation (S2TT) in the medical domain presents significant challenges due to the complexity of medical terminology and the scarcity of high-quality multilingual data. To address these issues, we propose MMST (Multilingual Medical Speech-to-text Translation), a novel framework that integrates domain knowledge into S2TT. Initially, we develop a multilingual medical terminology dictionary utilizing a large language model to extract terms from multilingual medical corpora. Following this, we create MMST, which consists of two main components: a two-stage training approach that integrates general pretraining with fine-tuning specific to the medical domain, and a terminology injection mechanism that embeds target terms into translation prompts, directing the generation process during training. Experiments on a many-to-many multilingual medical dataset demonstrate that MMST consistently outperforms the strong QwenAudio baseline, achieving an average BLEU improvement of +6.97 and higher BERTScore, particularly on terminology-rich and low-resource language pairs.