MOSS-Transcribe-Diarize 0.9B

MOSS-Transcribe-Diarize 0.9B

Transcribes long recordings with several speakers in one pass, adding timestamps and anonymous speaker labels, in 50+ languages. For teams turning meetings, calls, podcasts and lectures into readable transcripts.

Made by hf-audio on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MOSS-Transcribe-Diarize 0.9B is an audio model that writes down who said what, and when, for long recordings with more than one speaker. It does transcription and speaker labelling together, so you do not chain two separate tools. This copy is the Transformers-ready version hosted by hf-audio.

Who it suits: It suits anyone who handles long, messy recordings and wants a single transcript with timestamps and speaker tags, without building a pipeline.

Industry | Who uses it | How they use it | Example
Business | Teams that hold meetings and calls | Turn recordings into transcripts that mark who spoke | Minutes from a team call
Media | Podcasters and journalists | Transcribe interviews with the speakers separated | An interview with two voices
Education | Lecturers and students | Get timestamped text from recorded lectures | A transcript of a long lecture
Video | Subtitle makers | Produce subtitles from video with speaker labels | A subtitle file for a film clip

Skip it if you need real names for speakers, because the labels it assigns are anonymous.

Where to use it:

• Python with the transformers library, loading the model with remote code
• vLLM or SGLang, which the card lists for serving
• The card's subtitle web app, with an interface in Simplified Chinese and English

Use case | Who | What they get back
Meeting transcripts | Office teams | Timestamped text with speaker tags
Podcast and interview notes | Producers | A written record that separates each voice
Subtitle drafts | Video editors | Time-aligned text from an audio or video file
Term-heavy recordings | Specialists | Better handling of domain words you supply as hotwords

How to set it up: You need these first.

• Python with pip
• The transformers library, since the repository ships its own loading code
• Room for the weights: about 1.8 GB in BF16, or about 454 MB at 4-bit by estimate

1. Install the Hugging Face hub library with pip.
2. Download the model files with the hf command.
3. Load the model with the transformers library, allowing its remote code to run.
4. Follow the card's quickstart sections for Python, serving or the subtitle app.

pip install -U huggingface_hub
hf download hf-audio/MOSS-Transcribe-Diarize-HF

How to use it: Give it an audio or video file and it returns one speaker-aware transcript with timestamps. Speakers appear as tags like [S01] and [S02]. You can add your own instructions and hotwords for terms it should expect.

hf download hf-audio/MOSS-Transcribe-Diarize-HF

Tool | What it does
Transformers remote code | Loads the model when you allow remote code
vLLM and SGLang | Serving options the card names
Subtitle web app | A web interface for making subtitles
Hotword prompts | Steer the transcript toward terms you supply

Example: Give it a recording of a team meeting. You get one transcript where each line carries a time range and a speaker tag such as [S01] or [S02], so you can see who said what without a second tool.

For newcomers and experts: If you are new, think of it as a note taker that also tells voices apart. A first small try is a short recording with two people talking, to see how the speaker tags look.

If you know the field, the model does speech recognition and diarization jointly, not as separate stages. The card reports that it handles recordings up to 90 minutes in a single pass and can emit acoustic event notes. It evaluates with CER, cpCER and delta-cp on AISHELL-4, Alimeeting, Podcast and Movies. It also reports first place in a multilingual speech challenge.

Audio or video file in
One pass over the whole recording
Speech and speakers handled together
Timestamps and speaker tags added
Single transcript out

Good to know:

• Languages: the card says 50+, and the record lists en and zh
• Size: 909M parameters in Safetensors
• Remote code: the repository includes loading code that runs when you allow remote code
• Labels: speakers are anonymous tags, not names

• License: Apache 2.0
• Cost to run: none named, the source lists no API key, account or paid service
• Data it reaches: the audio or video files you give it, handled on your own machine, and the source names no web or account access
• Not for: naming speakers, since the card describes only anonymous labels

On Hugging Face, 13 November 2026

• Task: audio text to text
• Library: transformers
• Parameters: 909M
• Weights: about 1.8 GB in BF16; about 454 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 759
• Likes: 5
• License: apache-2.0
• Languages: en, zh
• Tags: transformers, safetensors, moss_transcribe_diarize, text2text-generation, moss, audio, speech, diarization, timestamp-asr, long-form-audio, multimodal, multilingual

Links

• Model: https://huggingface.co/hf-audio/MOSS-Transcribe-Diarize-HF
• Files: https://huggingface.co/hf-audio/MOSS-Transcribe-Diarize-HF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hf-audio/MOSS-Transcribe-Diarize-HF

Then follow the usage in its model card. Nothing in this kit runs by itself.