EMNLP 2026 · retrieval-augmented streaming speech translation
Four moments from real lectures. Both systems hear the same audio, 1.92 seconds at a time, and RASST also looks up a glossary for every chunk. The text below is the recorded output from the paper's evaluation runs, shown at the moment each system produced it.
Each clip is cut from a full talk that both systems translated end to end during evaluation, at latency multiplier 2 (1.92-second chunks, about two seconds behind the speaker). Text appears at the point in the audio where the system produced it in that run, and the lookup log shows the glossary entries RASST retrieved for each chunk, how long the lookup took, and whether the translation then used the entry. Nothing here runs live.
A term counts as right when the glossary's translation appears in the system's output for the clip, the same exact-form rule as the paper's term accuracy. The counts describe these clips only. Across full test sets, RASST gets 81% of oncology terms and 86% of NLP-talk terms right at this latency. Audio: .