MOSS-Transcribe-Diarize

MOSS-Transcribe-Diarize

Turns long recordings of several speakers into one timestamped transcript with anonymous speaker labels in a single pass, for teams that transcribe meetings, calls, podcasts and interviews in 50+ languages.

Made by OpenMOSS-Team 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 transcribes long recordings and labels who is speaking in the same pass. You do not need a separate speaker-detection tool. It returns text with timestamps and anonymous speaker tags, and it can handle recordings up to 90 minutes long in 50+ languages.

Who it suits: It suits teams that turn long, messy multi-speaker audio into readable transcripts, and developers who add that to their own tools.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Business | Operations and admin teams | Transcribe meetings and calls | Get a transcript of a team call with each speaker tagged
Media | Podcast and video producers | Produce subtitles and show notes | Turn an interview into timestamped text
Education | Schools and trainers | Transcribe lectures | Turn a recorded lecture into searchable text
Software | Developers | Add speaker-aware transcription to an app | Feed in a recording and read back labelled turns

If your recordings have one speaker and you do not need timestamps or speaker tags, a simpler transcription model may do.

Where to use it:

• Python scripts that use the transformers library
• vLLM or SGLang, which the card lists for serving the model
• A subtitle web app that comes with the project
• Windows, macOS or Linux with Python, for downloading the files

Title: Jobs it does
Use case | Who | What they get back
Meeting transcripts | Office teams | Timestamped text with a tag for each speaker
Subtitles | Video teams | Time-aligned text for captions
Term-heavy audio | Specialist teams | Better handling of domain terms through hotword prompts
Sound event notes | Analysts | Annotations of non-speech sounds in the audio

How to set it up: You need Python and about 1.8 GB of disk space for the BF16 weights, or about 454 MB for a 4-bit copy by Hugging Face's estimate.

• Python with pip
• The transformers library
• The huggingface_hub package

1. Download the files with the two Hugging Face commands.
2. Load the model with transformers. The repository ships custom code, so loading needs trust_remote_code turned on.
3. For environment setup, serving or the subtitle app, follow the card's Quickstart sections.

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-Transcribe-Diarize

How to use it: Give the model an audio or video file and it returns a compact transcript in one pass. Each turn carries a timestamp and a speaker tag such as S01 or S02. You can add your own instructions and a list of hotwords for names or terms it might otherwise miss.

Example: You pass in a 60-minute team meeting and add the names of two products as hotwords. The transcript comes back with timestamps, with each turn tagged to speaker S01, S02 or another, and the product names spelled the way you listed them.

For newcomers and experts: Think of it as a note-taker that writes down what was said and who said it. A first small test is a short two-person recording. The labels are anonymous, so you map them to real names yourself.

The model replaces a pipeline of separate speech recognition and diarization systems with one model that does both jointly. It scores itself with character error rate, concatenated minimum-permutation character error rate and a delta between the two, and lower is better for all three. The card says it took first place at the 2nd MLC-SLM Challenge at INTERSPEECH 2026.

Audio or video file goes in
One model transcribes and tracks speakers
Timestamps and speaker tags are added
Speaker-aware transcript comes out

Good to know: The card lists a subtitle web app that supports Simplified Chinese and English. Other people have published 16 quantized and 34 fine-tuned versions. The speaker labels are anonymous and do not name anyone.

• License: Apache 2.0.
• Cost to run: none named.
• Data it reaches: the files come from Hugging Face, and the audio you transcribe stays on the machine where you run the model.
• Not for: naming speakers, since the labels are anonymous.

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: 121,031
• Likes: 520
• License: apache-2.0
• Languages: en, zh
• Quantized versions: 16
• Fine-tunes: 34
• Tags: transformers, safetensors, moss_transcribe_diarize, text2text-generation, moss, audio, speech, diarization, timestamp-asr, long-form-audio, multimodal, multilingual

Links

• Model: https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize
• Files: https://huggingface.co/OpenMOSS-Team/MOSS-Transcribe-Diarize/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-Transcribe-Diarize

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