Fused Acoustic and Text Encoding for Multimodal Bilingual Pretraining and Speech Translation

Recently, representation learning for text and speech has successfully\nimproved many language related tasks. However, all existing methods suffer from\ntwo limitations: (a) they only learn from one input modality, while a unified\nrepresentation for both speech and text is needed by tasks such as end-to-end\nspeech translation, and as a result,(b) they can not exploit various\nlarge-scale text and speech data and their performance is limited by the\nscarcity of parallel speech translation data.To address these problems, we\npropose a Fused Acoustic and Text Masked Language Model (FAT-MLM) which jointly\nlearns a unified representation for both acoustic and text input from various\ntypes of corpora including parallel data for speech recognition and machine\ntranslation, and even pure speech and text data. Within this cross-modal\nrepresentation learning framework, we further present an end-to-end model for\nFused Acoustic and Text Speech Translation (FAT-ST). Experiments on three\ntranslation directions show that by fine-tuning from FAT-MLM, our proposed\nspeech translation models substantially improve translation quality by up to\n+5.9 BLEU.\n

Paper

Similar papers

© 2026 NYSGPT2525 LLC