Forward Translation to Mix Data for Speech Translation

End-to-End speech translation means that using a model to translate speech in one language into text in another language. Currently, the main challenge in the field of speech translation is data scarcity. Existing works solve this problem by using text information or applying data augmentation. However, these works only focus on the exploitation of a single corpus, ignoring the full use of existing human-labeled different-sources data. In this paper, we introduce a simple method to solve the data scarcity problem: training a model with simply mixed data and applying the forward translation method to expand the training set. We perform experiments on covost v2 French-English and mTEDx French-English. Our experiments demonstrate that combining the mixture of speech translation corpora with forward translation can yield a better result than the method without mixing.