Direct Speech-to-Text Translation Models as Students of Text-to-Text Models

Direct speech-to-text translation (ST) is an emerging approach that consists in performing the ST task with a single neural model. Although this paradigm comes with the promise to outperform the traditional pipeline systems, its rise is still limited by the paucity of speech-translation paired corpora compared to the large amount of speech-transcript and parallel bilingual corpora available to train previous solutions. As such, the research community focused on techniques to transfer knowledge from automatic speech recognition (ASR) and machine translation (MT) models trained on huge datasets. In this paper, we extend and integrate our recent work (Gaido, Gangi, et al. 2020) analysing the best performing approach to transfer learning from MT, which is represented by knowledge distillation (KD) in sequence-to-sequence models. After the comparison of the different KD methods to understand which one is the most effective, we extend our previous analysis of the effects – both in terms of benefits and drawbacks – to different language pairs in high-resource conditions, ensuring the generalisability of our findings. Altogether, these extensions complement and complete our investigation on KD for speech translation leading to the following overall findings: i) the best training recipe involves a word-level KD training followed by a fine-tuning step on the ST task, ii) word-level KD from MT can be detrimental for gender translation and can lead to output truncation (though these problems are alleviated by the fine-tuning on the ST task), and iii) the quality of the ST student model strongly depends on the quality of the MT teacher model, although the correlation is not linear.