What is MOSS-Transcribe-Diarize 0.9B?
MOSS-Transcribe-Diarize is an open speech-to-text model from the OpenMOSS team, open-sourced on 9 July 2026 under Apache 2.0. Instead of chaining Whisper with a separate diarisation model such as pyannote, it emits one stream of timestamped, speaker-labelled text in the form [start][S01]text[end], with optional acoustic event tags and custom hotwords.
It pairs a Whisper-Medium-style audio encoder with a Qwen3-0.6B-style decoder, about 0.9 billion parameters in total. The model card claims 50+ languages without a language setting and single-pass inference on recordings up to 90 minutes. A larger MOSS-Transcribe-Diarize Pro is available only through MOSI's online playground.
The team reports first place in the 2nd MLC-SLM Challenge at INTERSPEECH 2026, which covered 14 languages. The weights have about 170,000 downloads.
Is MOSS-Transcribe-Diarize 0.9B worth it?
MOSS-Transcribe-Diarize does in one 0.9B model what most open stacks need two for: who spoke, when and what they said. For a team that has been gluing Whisper to pyannote and reconciling their timestamps, that removes a whole class of alignment bugs.
The published results are strong but narrow. On the AISHELL-4 meeting set the team reports a speaker-attributed error rate (cpCER) of 15.83, against 37.95 for ElevenLabs, 27.43 for Gemini 3 Pro and 24.99 for Microsoft's VibeVoice ASR; on its Podcast set the figure is 7.37 against ElevenLabs' 11.34. These are self-reported character error rates, and AISHELL-4 and AliMeeting are Mandarin meeting corpora, so they say little about English call-centre audio.
Throughput on one H100 is practical. In the SGLang Omni cookbook figures, long recordings run at about 51 seconds of audio per second with one request and about 99 with 16 in parallel, so one GPU clears an hour-long meeting in well under a minute at full load. Short clips return in about 0.39 seconds.
The rough edges are operational. Loading needs trust_remote_code, vLLM support requires a pinned nightly build, long files need max_new_tokens raised from 5,120 to as much as 65,536, and the eager attention fallback runs out of memory on long recordings.
Where it earns its place
A podcast network publishing 200 episodes a month
Produce speaker-labelled transcripts and subtitles
Run the model behind vLLM on one GPU and feed whole episodes up to 90 minutes; the bundled subtitle web app exports captions. No per-minute fees, against paying a hosted diarisation API.
An ML engineer at a Shanghai meeting-software company
Replace a Whisper plus pyannote pipeline for Mandarin meetings
One model call returns [start][Sxx]text[end] segments, which removes timestamp reconciliation; the team's AISHELL-4 cpCER of 15.83 is the most relevant published number for this job.
A legal-tech startup handling depositions that must stay on-premise
Transcribe two-hour recordings with speaker turns
Files beyond 90 minutes must be split, and English accuracy is not well documented, so run a pilot against a human transcript before committing.
Before you commit
- Setup
- Clone the repo, create a Python 3.12 environment with uv and install with the torch-runtime extra; load with Transformers and trust_remote_code=True. Optional flash-attn speeds up CUDA inference.
- Integrations
- Transformers, SGLang Omni, vLLM (OpenAI-compatible /v1/audio/transcriptions) and the FunASR ecosystem; a fine-tuning guide and a subtitle web app are included.
- Data and privacy
- Apache 2.0 weights with no revenue cap; self-hosted audio stays on your hardware. The Pro playground is a MOSI service whose retention terms we could not find.
- Cost at scale
- One H100 handles roughly 51 to 99 seconds of long-form audio per second depending on concurrency, by the cookbook figures, so a month of 10,000 audio hours needs roughly 100 to 200 H100 hours.
- Limits
- 90 minutes per pass; default max_new_tokens of 5,120 truncates long transcripts; eager attention runs out of memory on long audio; English and European-language accuracy are not published in detail.
What's good
- Transcription, diarisation and timestamps in one pass
- Up to 90 minutes per file without segmentation
- Apache 2.0 weights at 0.9B parameters
- Serves through vLLM's OpenAI-compatible transcription API
What's not
- Headline benchmarks are self-reported and mostly Mandarin
- Requires trust_remote_code and a pinned vLLM nightly
- Long files need large max_new_tokens settings
- Best model (Pro) is playground-only
Best for
Teams transcribing long multi-speaker recordings on their own GPUs, especially in Chinese, who want speaker labels without a separate diarisation model.
MOSS-Transcribe-Diarize 0.9B alternatives
MOSS-Transcribe-Diarize 0.9B: common questions
- Is MOSS-Transcribe-Diarize 0.9B worth it?
- MOSS-Transcribe-Diarize replaces a two-model meeting pipeline with one small one, and its benchmark wins are mostly in Mandarin. We rate it 3.9 out of 5.
- How much does MOSS-Transcribe-Diarize 0.9B cost?
- Free under Apache 2.0. The larger Pro model is offered through MOSI's online playground; API pricing is not published.
- What are the downsides of MOSS-Transcribe-Diarize 0.9B?
- Headline benchmarks are self-reported and mostly Mandarin; requires trust_remote_code and a pinned vLLM nightly; long files need large max_new_tokens settings; best model (Pro) is playground-only.
- Who is MOSS-Transcribe-Diarize 0.9B best for?
- Teams transcribing long multi-speaker recordings on their own GPUs, especially in Chinese, who want speaker labels without a separate diarisation model.
- How much does MOSS-Transcribe-Diarize 0.9B cost for a team or in production?
- One H100 handles roughly 51 to 99 seconds of long-form audio per second depending on concurrency, by the cookbook figures, so a month of 10,000 audio hours needs roughly 100 to 200 H100 hours.
- Is MOSS-Transcribe-Diarize 0.9B safe for company data?
- Apache 2.0 weights with no revenue cap; self-hosted audio stays on your hardware. The Pro playground is a MOSI service whose retention terms we could not find.
- What are the best MOSS-Transcribe-Diarize 0.9B alternatives?
- The closest alternatives we have reviewed are VibeVoice-ASR, Whisper large-v3-turbo, ElevenLabs Scribe v2 and Parakeet TDT 0.6B v3.
Answers checked Oct 26, 2026.