Harnessing Indirect Training Data for End-to-End Automatic Speech Translation: Tricks of the Trade
This work addresses the problem of limited training data for end-to-end automatic speech translation, which is incremental as it builds on existing methods to improve performance.
The paper tackled the performance gap between end-to-end and cascaded models in automatic speech translation by evaluating data augmentation and pretraining approaches, closing the gap from 8.2 to 1.4 BLEU on English-French and from 6.7 to 3.7 BLEU on English-Romanian datasets.
For automatic speech translation (AST), end-to-end approaches are outperformed by cascaded models that transcribe with automatic speech recognition (ASR), then translate with machine translation (MT). A major cause of the performance gap is that, while existing AST corpora are small, massive datasets exist for both the ASR and MT subsystems. In this work, we evaluate several data augmentation and pretraining approaches for AST, by comparing all on the same datasets. Simple data augmentation by translating ASR transcripts proves most effective on the English--French augmented LibriSpeech dataset, closing the performance gap from 8.2 to 1.4 BLEU, compared to a very strong cascade that could directly utilize copious ASR and MT data. The same end-to-end approach plus fine-tuning closes the gap on the English--Romanian MuST-C dataset from 6.7 to 3.7 BLEU. In addition to these results, we present practical recommendations for augmentation and pretraining approaches. Finally, we decrease the performance gap to 0.01 BLEU using a Transformer-based architecture.