Title: CUNI’s Submission to IWSLT 2026: A Compact Offline Model for Concurrent Speech Translation

Abstract: For the IWSLT 2026 Simultaneous Speech Translation Shared task, we present a submission covering Czech-to-English and English-to-German/Italian translation directions. Our approach integrates the state-of-the-art AlignAtt policy into the offline direct speech-to-text translation framework known as Canary to achieve simultaneous translation capabilities. The proposed system offers three primary advantages: (1) superior translation accuracy, surpassing comparable baseline models in both low- and high-latency scenarios during computationally unconstrained simulations; (2) efficiency, driven by a compact architecture with merely 1 billion parameters; and (3) extensive multilingual support, enabling interaction across 25 source and 25 target languages.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC