Large-scale pre-training has been shown to benefit speech translation tasks. However, existing multimodal pre-training efforts rely on parallel corpora for semantic alignment, potentially limiting ...