Textless Speech-to-Speech Translation With Limited Parallel Data
This addresses the challenge of speech-to-speech translation for languages lacking text resources or with scarce parallel data, though it is incremental as it builds on existing pretraining and backtranslation techniques.
The paper tackles the problem of speech-to-speech translation for textless languages or language pairs with limited parallel data by proposing PFB, a framework that requires only dozens of hours of parallel speech data, achieving performance within 1-2 points of higher-resourced baselines on domains like European Parliament and Common Voice.
Existing speech-to-speech translation (S2ST) models fall into two camps: they either leverage text as an intermediate step or require hundreds of hours of parallel speech data. Both approaches are incompatible with textless languages or language pairs with limited parallel data. We present PFB, a framework for training textless S2ST models that require just dozens of hours of parallel speech data. We first pretrain a model on large-scale monolingual speech data, finetune it with a small amount of parallel speech data (20-60 hours), and lastly train with an unsupervised backtranslation objective. We train and evaluate our models for English-to-German, German-to-English and Marathi-to-English translation on three different domains (European Parliament, Common Voice, and All India Radio) with single-speaker synthesized speech. Evaluated using the ASR-BLEU metric, our models achieve reasonable performance on all three domains, with some being within 1-2 points of our higher-resourced topline.