MOSS-Transcribe-Diarize-GGUF

MOSS-Transcribe-Diarize-GGUF

Turns audio into timestamped transcripts with speaker labels in one pass. A GGUF conversion of a 0.9B model for CrispASR, for people who want meeting or interview transcripts on their own computer.

Made by cstr 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: This is a GGUF copy of OpenMOSS-Team's MOSS-Transcribe-Diarize, prepared by cstr for the CrispASR program. One 0.9B model does speech recognition, tells speakers apart and adds timestamps. You get a transcript where each line has a time range and a speaker label. It supports English and Chinese.

Who it suits: People who need written transcripts of recordings with the speakers marked, and who prefer to run the model locally.

Industry | Who uses it | How they use it | Example
Journalism | Reporters with interview recordings | Run the model over an audio file | Getting a transcript that shows which speaker said each line
Business | Teams that record meetings | Transcribe a meeting file with a list of names and terms | Adding product and team names to the hotwords list
Research | Researchers who collect spoken data | Convert recordings to timestamped text | Producing subtitle style output for later review

Skip it if you need a hosted transcription service, or if your audio is in a language other than English or Chinese.

Where to use it: It is built for CrispASR, a command line program, and the files download on Windows, macOS or Linux.

• CrispASR: the program that loads the GGUF files
• Windows, macOS or Linux: where you can fetch the files with Python
• A terminal: where you run the transcription command

Use case | Who | What they get back
Meeting transcripts | Teams | Text with times and speaker labels
Interview notes | Journalists | A labelled transcript to check against the audio
Subtitle files | Video editors | Timed text from a recording

How to set it up: You need Python to download the files and CrispASR to run them.

• Python with pip
• CrispASR installed on your computer
• Disk space for one model file, between about 1.2 GB and 1.7 GB

1. Install the Hugging Face hub library.
2. Download the model files.
3. Pick a file size and point CrispASR at it.

pip install -U huggingface_hub
hf download cstr/MOSS-Transcribe-Diarize-GGUF

The repository holds three sizes of the model.

File type | Size | What it is
F16 | 1.7 GB | Full precision copy
Q8_0 | 1.4 GB | Smaller copy at 8 bits
Q4_K | 1.2 GB | Smallest copy at 4 bits, the one its author recommends

How to use it: Run crispasr with the moss-diarize backend on an audio file. The first command lets the program pick the model file by itself.

crispasr --backend moss-diarize -m auto -f audio.wav -osrt

To give it a list of words it should expect, such as names or product terms:

crispasr --backend moss-diarize \
  -m moss-transcribe-diarize-0.9b-q4_k.gguf \
  -f meeting.wav \
  --hotwords "MOSS,CrispASR" \
  -osrt

Tool | What it does
crispasr | Runs the transcription with the moss-diarize backend
-m | Chooses the model file, or auto
-f | Names the audio file to transcribe
--hotwords | Passes a comma separated list of words to expect

Example: You point CrispASR at a recording called meeting.wav, with the hotwords MOSS and CrispASR. It returns a transcript in which every line carries a start and end time and a speaker label such as Speaker 1 or Speaker 2.

For newcomers and experts: If you are new to this, record a short clip with two voices and run the first command on it. Check that the output names two speakers and that the times line up with the audio.

If you know the field, the card describes a Whisper style audio encoder with a temporal merge and an adapter, feeding a Qwen3-0.6B decoder. The decoder writes timestamp and speaker tokens along with the text. This repository only converts the format and quantizes it, with no retraining.

Audio file goes into CrispASR
Whisper style encoder reads the audio
Qwen3 decoder writes tokens
Timed, speaker labelled text comes out

Good to know: The authors state that this is a conversion only, so the original model decides what it can and cannot do.

• License: Apache 2.0, the same as the base model
• Cost to run: none named
• Data it reaches: the audio files you point it at
• Not for: new training, since no training or fine tuning was done here. Questions about training data go to the original authors

On Hugging Face, 13 November 2026

• Parameters: 909M
• Weights: about 454 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 2,616
• Likes: 10
• License: apache-2.0
• Languages: en, zh
• Based on: OpenMOSS-Team/MOSS-Transcribe-Diarize
• Tags: gguf, speech-recognition, speaker-diarization, timestamps, crispasr, multilingual

Links

• Model: https://huggingface.co/cstr/MOSS-Transcribe-Diarize-GGUF
• Files: https://huggingface.co/cstr/MOSS-Transcribe-Diarize-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download cstr/MOSS-Transcribe-Diarize-GGUF

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