This repository contains the already-converted Apple Silicon MLX model. Users do not need to download the original HF model or run conversion locally.

Use

Install the project:

bash

git clone https://github.com/OpenMOSS/MOSS-Transcribe-Diarize.git mlx-MOSS-Transcribe-Diarize
cd mlx-MOSS-Transcribe-Diarize
python -m pip install -e ".[mlx-runtime]"

Run transcription:

bash

python -m moss_transcribe_diarize.mlx.cli /path/to/input.wav \
  --model vanch007/mlx-MOSS-Transcribe-Diarize \
  --out-dir runs/mlx_example

Python:

python

from moss_transcribe_diarize.mlx import load_model

model = load_model("vanch007/mlx-MOSS-Transcribe-Diarize", strict=True)
result = model.generate("/path/to/input.wav", max_tokens=2048, temperature=0.0)
print(result.text)

Output Format

[start_time][Sxx]transcribed speech[end_time]

Example:

[0.06][S01] Hello world. This is a local MLX smoke test.[3.12]

Conversion

Converted from OpenMOSS-Team/MOSS-Transcribe-Diarize with:

bash

python -m moss_transcribe_diarize.mlx.convert \
  --source pretrained/moss-transcribe-diarize-hf \
  --output pretrained/mlx-moss-transcribe-diarize \
  --overwrite

See mlx_conversion.json for conversion metadata.