Multilingual Speech Translation with Efficient Finetuning of Pretrained Models

Abstract

We present a simple yet effective approach to build multilingual speech-to-text (ST) translation by efficient transfer learning from pretrained speech encoder and text decoder. Our key finding is that a minimalistic LNA (LayerNorm and Attention) finetuning can achieve zero-shot crosslingual and cross-modality transfer ability by only finetuning less than 10% of the pretrained parameters. This enables effectively leveraging large pretrained models with low training cost. Using wav2vec 2.0 for acoustic modeling, and mBART for multilingual text generation, our approach advanced the new state-of-the-art for 34 translation directions (and surpassing cascaded ST for 23 of them) on large-scale multilingual ST benchmark CoVoST 2 (+6.4 BLEU on average across 15 En-X directions and +5.1 BLEU on average across 19 X-En directions). Our approach demonstrates strong zero-shot performance in a many-to-many multilingual model (+5.7 BLEU on average across 18 non-English directions), making it an appealing approach for attaining high-quality speech translation with improved parameter and data efficiency.

References in corpus (15)

• Cross-lingual Language Model Pretraining
• VL-BERT: Pre-training of Generic Visual-Linguistic Representations
• Multilingual Denoising Pre-training for Neural Machine Translation
• vq-wav2vec: Self-Supervised Learning of Discrete Speech Representations
• Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges
• Multilingual Translation with Extensible Multilingual Pretraining and Finetuning
• Parameter-Efficient Transfer Learning for NLP
• Listen and Translate: A Proof of Concept for End-to-End Speech-to-Text Translation
• fairseq S2T: Fast Speech-to-Text Modeling with fairseq
• Beto, Bentz, Becas: The Surprising Cross-Lingual Effectiveness of BERT
• Pareto Multi-Task Learning
• CoVoST 2 and Massively Multilingual Speech-to-Text Translation
• End-to-End Speech Translation with Knowledge Distillation
• Recipes for Adapting Pre-trained Monolingual and Multilingual Models to Machine Translation
• What Would Elsa Do? Freezing Layers During Transformer Fine-Tuning

Cited by in corpus (6)

• End-to-end Speech Translation via Cross-modal Progressive Training
• Large-Scale Self- and Semi-Supervised Learning for Speech Translation
• Pay Better Attention to Attention: Head Selection in Multilingual and Multi-Domain Sequence Modeling
• fairseq S^2: A Scalable and Integrable Speech Synthesis Toolkit
• Improving Speech Translation by Understanding and Learning from the Auxiliary Text Translation Task
• FST: the FAIR Speech Translation System for the IWSLT21 Multilingual Shared Task