Regularizing End-to-End Speech Translation with Triangular Decomposition Agreement

End-to-end speech-to-text translation (E2E-ST) is becoming increasingly popular due to the potential of its less error propagation, lower latency, and fewer parameters. Given triplet training corpus〈speech, transcription, translation〉, conventional high-quality E2E-ST system leverages the〈speech, transcription〉pair pre-train model then utilizes translation〉pair optimize it further. However, this process only involves two-tuple data at each stage, loose coupling fails fully exploit association between data. In paper, we attempt joint probability transcription based on speech input directly leverage such Based that, propose a novel regularization method for improve agreement dual-path decomposition within data, which should be equal in theory. To achieve goal, introduce two Kullback-Leibler divergence terms into objective reduce mismatch output probabilities dual-path. Then well-trained can naturally transformed as models by pre-defined early stop tag. Experiments MuST-C benchmark demonstrate that our proposed approach significantly outperforms state-of-the-art baselines all 8 language pairs while achieving better performance automatic recognition task.