Slide 1
A point-nine-billion-parameter open model takes long, multi-speaker audio and returns a transcript that already knows who said what. That's MOSS-Transcribe-Diarize, and it does the whole thing in a single pass.
One pass in, a speaker-labeled transcript out.
Feed it a meeting or a podcast, up to ninety minutes, in fifty or more languages. Back comes one stream: start and end times, a speaker tag like S-oh-one, and the words, all interleaved.
No separate diarizer — who-spoke falls out of the same stream.
Normally this is two systems: one to transcribe, a second to work out who spoke and stitch it back together. Here it's one. A trainable audio encoder feeds a single speech model that emits the timestamps, the speaker, and the text directly, so who-spoke falls out of the same stream.
Lowest cpCER of any non-Pro system, all four suites
On OpenMOSS's own benchmark, the result is the surprise. Take AISHELL-4, a hard Mandarin meeting set. The metric is cpCER: character error once you also have to get who-spoke right, lower is better. MOSS scores fifteen point eight. Gemini 3 Pro, on the same eval, twenty-seven point four. Across all four suites the open model has the lowest speaker-attributed error of any non-Pro system, and on podcasts its diarization penalty is the smallest in the column, under even the closed Pro tier.
1st place on the transcription task — not the lab's own scoreboard.
One result here isn't the lab's own. At the second MLC-SLM challenge, part of Interspeech, MOSS took first place on the transcription and diarization task, across fourteen languages, on an independent leaderboard.
The honest catch: whose numbers, and the one cell it loses.
Now the honest part. Almost every number here is OpenMOSS's own eval, run on their own long-form protocol, one the report admits GPT-4o and Gemini can't reliably handle, so beating them means on this task, not general speech recognition. There's also a stronger closed Pro tier that tops most columns. And it doesn't win everything: on overlapping movie speech, its diarization still trails Gemini 3 Pro.
Slide 7
github.com/OpenMOSS/MOSS-Transcribe-Diarize · arXiv 2601.01554 · Apache-2.0
So a job that used to need two systems now runs from one small open model. Weights, code, and a local subtitle app, all Apache-2.0.
