Figure 6: BLEU ontestsetachievedbyS2Tmodels trained onCoVoSTtrainset+mined dataat differentthresholds. 15

A.1 SpeechTranslationEvaluation One hyperparameter in our speech translation evaluation is the threshold on the alignment scores. Speech translation models are trained on the combination of CoVoST2 train set andmined dataatdifferent thresholds. Based on Figure 5, the optimal threshold ist = 1.07for de-en, es-en and fr-en directions. As we decrease the threshold, more mined data is added to the train set improving model performance. For the completeness of our analysis, we also report BLEU scores on Common Voice test set in Figure6.