Dubformer today announced its sponsorship of the WMT26 General Machine Translation shared task, with CTO and co-founder Sergey Dukanov joining the General MT organizing team to lead work on Spoken Dialogue, a new domain that evaluates translation with original video and machine-generated transcripts.

WMT26 is the Eleventh Conference on Machine Translation, co-located with EMNLP 2026 in Budapest in November 2026. Dubformer's sponsorship covers test-set translations and human evaluation costs, supporting the shared comparison that the General MT task depends on.

The Spoken Dialogue domain brings that problem into the General MT setting. Systems receive machine-generated transcripts, and human evaluators use the original video when judging outputs. That matters for translated speech, where timing and what happens on screen can change whether a line works.

"Current video translation systems treat the task as text translation with an ASR front-end — quietly cutting off a lot of useful signal. ASR is noisy, and even a clean transcript is a lossy projection: the visuals carry the factual disambiguation that words leave open (gestures, on-screen objects, gender, who is being addressed), while the audio carries an affective layer that text strips away (prosody, emotion, intent). Both are needed. Our WMT26 speech translation test set concentrates exactly the segments where these channels matter, so it measures how far current systems still are from a human translator working with the full video." said Sergey Dukanov, Dubformer's CTO and co-founder.

Dubformer at WMT

Dubformer's involvement with WMT predates this sponsorship. Dvorkovich has co-authored WMT findings papers since 2022, and Dvorkovich and Dukanov co-authored the WMT25 findings paper on evaluating beyond easy test sets. At WMT24, the findings paper noted that a proprietary Dubformer engine was used to prepare English-language speech material for the task. Dubformer's General MT submission that year ranked among the top systems across five language pairs and placed first among machine translation systems for English-to-Spanish and English-to-Russian on the speech-domain tests.

WMT26 General MT details

• Test data release: 18 June 2026

• Translation submission deadline: 2 July 2026

• Conference: November 2026 in Budapest, co-located with EMNLP 2026

The General MT task includes document-level test sets and instruction-following context. Human evaluators score primary submissions, and the Spoken Dialogue domain is judged with reference to the original video. The full task description is available on the WMT26 General MT page.