Lightweight Adapter Tuning for Multilingual Speech Translation

• DOI: 10.18653/v1/2021.acl-short.103
• Publication Date: 2021-07-27T01:42:51Z
• Authors: Hang Le; Juan Pino; Changhan Wang; Jiatao Gu; Didier Schwab; Laurent Besacier
• Keywords: Adapter (computing); Speech translation

Adapter modules were recently introduced as an efficient alternative to fine-tuning in NLP. tuning consists freezing pretrained parameters of a model and injecting lightweight between layers, resulting the addition only small number task-specific trainable parameters. While adapter was investigated for multilingual neural machine translation, this paper proposes comprehensive analysis adapters speech translation (ST). Starting from different pre-trained models (a ST trained on parallel data or BART (mBART) non-parallel data), we show that can be used to: (a) efficiently specialize specific language pairs with low extra cost terms parameters, (b) transfer automatic recognition (ASR) task mBART task. Experiments offer competitive results full fine-tuning, while being much more parameter-efficient.