ASR Posterior-based Loss for Multi-task End-to-end Speech Translation

End-to-end speech translation (ST) translates source language speech directly into target language without an intermediate automatic speech recognition (ASR) output, as in a cascading approach. End-to-end ST has the advantage of avoiding error propagation from the intermediate ASR results, but its performance still lags behind the cascading approach. A recent effort to increase performance is multi-task learning using an auxiliary task of ASR. However, previous multi-task learning for end-to-end ST using cross entropy (CE) loss in ASR-task targets one-hot references and does not consider ASR confusion. In this study, we propose a novel end-to-end ST training method using ASR loss against ASR posterior distributions given by a pre-trained model, which we call ASR posterior-based loss. The proposed method is expected to consider possible ASR confusion due to competing hypotheses with similar pronunciations. The proposed method demonstrated better BLEU results in our Fisher Spanish-to-English translation experiments than the baseline with standard CE loss with label smoothing.