MOSS-Transcribe-Diarize
OpenMOSS's 0.9B Apache 2.0 model that turns up to 90 minutes of multi-speaker audio into a speaker-labeled, time-stamped transcript.
In short. MOSS-Transcribe-Diarize listens to a long meeting, call or podcast and writes down who said what and when, in one pass. It tells speakers apart with labels like S01 and S02, accepts a list of special terms to spell correctly, and is small enough to run on a single GPU.
Summary
MOSS-Transcribe-Diarize 0.9B is an end-to-end model that turns a long, multi-speaker recording into one speaker-labeled, time-stamped transcript, instead of chaining a speech recognizer to a separate diarization system. According to its card, it handles recordings of up to 90 minutes in a single pass, transcribes and separates speakers in more than 50 languages, accepts hotwords in the prompt and can mark acoustic events. The output is plain text in the form [start][S01]speech[end], and the speaker labels are anonymous and only meaningful within that recording. The configuration shows a Whisper-style audio encoder feeding a Qwen3-style 28-layer text decoder with 131,072 positions, about 0.91B parameters in all. The technical report ties the 128K window to the 90-minute limit.
On the card's tests, concatenated minimum-permutation character error rate (cpCER, lower is better) is 15.83% on AISHELL-4 meetings, 22.17% on Alimeeting, 7.37% on a podcast set and 12.76% on movie clips. VibeVoice-ASR scores 24.99%, 29.33%, 48.30% and 42.54% on the same sets, and the best listed commercial system scores 27.43% (Gemini 3 Pro), 32.84% (Gemini 3 Pro), 10.23% (Gemini 2.5 Pro) and 14.73% (Gemini 3 Pro). The sets are Chinese-language, the numbers are the developer's own, and the table also lists a stronger MOSS Transcribe Diarize Pro that this repository does not contain. The card also says the model won first place in the 2nd MLC-SLM Challenge at INTERSPEECH 2026, which covered 14 languages, but it gives no per-language scores for the rest of the 50-plus languages.
The license is Apache 2.0 and the weights are 1.82 GB in BF16. The repository needs custom Transformers code, loaded with trust_remote_code, and the card recommends serving through SGLang Omni, or a pinned vLLM nightly on CUDA 12, behind an OpenAI-style transcription endpoint. The GitHub package adds a local subtitle app that exports JSON, SRT and ASS files and can burn subtitles into video with FFmpeg. The card dates the release to 9 July 2026. In the 30 days to 11 November 2026, Hugging Face counted about 128,000 downloads.
Good for
- Meetings, interviews, calls and podcasts where you need who spoke and when from one model instead of a separate diarization pipeline
- Recordings up to 90 minutes long, with hotwords for names and domain terms
- Making subtitles for video, since the toolkit exports SRT and ASS files and can burn them in
Watch out for
- Speaker labels such as S01 are relative to one recording and do not identify real people, so matching voices across recordings needs another system
- The reported scores come from the developer, use Chinese-language sets, and include a stronger Pro model that is not in this repository, while the card gives no per-language table for the 50-plus languages
- It needs custom loading code with trust_remote_code and a CUDA setup, SGLang does not support CUDA 12 so the card points to a pinned vLLM nightly, and no minimum memory is stated
- Long recordings produce long outputs, and the card raises max_new_tokens to 65,536 for them, so a low limit will cut the transcript short
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of MOSS-Transcribe-Diarize. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| MOSS-Transcribe-Diarize 0.9B BF16 | 910M | 1.82 GB | 2.09 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is MOSS-Transcribe-Diarize?
- MOSS-Transcribe-Diarize listens to a long meeting, call or podcast and writes down who said what and when, in one pass. It tells speakers apart with labels like S01 and S02, accepts a list of special terms to spell correctly, and is small enough to run on a single GPU.
- Can I use MOSS-Transcribe-Diarize commercially?
- Yes. MOSS-Transcribe-Diarize is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is MOSS-Transcribe-Diarize open source?
- MOSS-Transcribe-Diarize is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run MOSS-Transcribe-Diarize on my own computer?
- Yes. The smallest variant, MOSS-Transcribe-Diarize 0.9B, needs about 2.09 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was MOSS-Transcribe-Diarize released?
- MOSS-Transcribe-Diarize was released by OpenMOSS on July 9, 2026.
- How long a context window does MOSS-Transcribe-Diarize have?
- MOSS-Transcribe-Diarize supports a context window of up to 128K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- MOSS-Transcribe-Diarize 0.9B model card (Hugging Face) huggingface.co
- MOSS-Transcribe-Diarize repository (GitHub) github.com
- MOSS Transcribe Diarize Technical Report (arXiv 2601.01554) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| VibeVoice-ASR Microsoft | 2.81B – 8.67B | 64K | MIT (Permissive) | 6.47 GB Laptop or phone | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Nemotron 3.5 Transcribe NVIDIA | 34B 4.3B active | 64K | OpenMDW-1.1 (Permissive) | 25 GB Gaming GPU | |
| Cohere Transcribe Cohere | 2.07B | – | Apache 2.0 (Permissive) | 4.75 GB Laptop or phone |