Omnilingual MT: Machine Translation for 1,600 Languages
Outstanding Paper Award at EMNLP 2026. We present Omnilingual MT, the first MT system benchmarked on more than 1,600 languages, where 1B–8B parameter specialized models match or …
Outstanding Paper Award at EMNLP 2026. We present Omnilingual MT, the first MT system benchmarked on more than 1,600 languages, where 1B–8B parameter specialized models match or …
We introduce BOUQuET, a multi-way, multicentric, multi-domain, paragraph-level dataset and benchmark for machine translation evaluation, designed for crowd-sourced expansion to …
We propose a character-based translation model to improve adaptability to new languages and modalities, particularly for low-resource scenarios. Our method achieves …
We propose MaskVAT, a generative Video-to-Audio (V2A) model combining a high-quality audio codec with a masked generative model to simultaneously achieve high audio quality, …
We investigate whether speech-to-text translation systems utilize prosody by introducing a new benchmark, ContraProSt. Our findings show that while models represent prosody …
We introduce ZeroSwot, a zero-shot speech translation method that aligns a speech encoder with a multilingual MT model using only ASR data, achieving state-of-the-art results …
We propose SegAugment, a data augmentation strategy that creates multiple sentence-level variations from document-level speech data, leading to significant performance gains in …
This paper details our submission to the IWSLT 2023 Speech Translation task, which utilizes wav2vec 2.0 and mBART50 foundation models. Our method incorporates a Siamese pretraining …
We present a new method to explain how Transformer models use context for language generation, demonstrating superior alignment with linguistic phenomena and shedding light on the …
We propose a Perceiver-based encoder with a novel Dynamic Latent Access (DLA) training method for efficient Speech Translation. This approach maps long speech inputs to a …
We propose Supervised Hybrid Audio Segmentation (SHAS), a method that learns optimal speech segmentation from manually segmented data. SHAS significantly improves translation …
Our submission to the IWSLT 2022 shared task details an end-to-end speech translation system built on large pretrained models. We leverage efficient fine-tuning techniques like …
Our submission to the WMT 2021 news translation shared task for German-French translation demonstrates that fine-tuning a pretrained mBART50 model significantly outperforms a …
Our submission to the IWSLT 2021 shared task details an end-to-end speech translation system combining large pretrained models with adapters for efficient fine-tuning. By training …