Property	Value
Base model	MediaTek-Research/Breeze-ASR-25
Architecture	Whisper Large (32 encoder + 32 decoder layers)
Parameters	~1.5B
Format	MLX safetensors (float16)
Size	~2.9 GB
Language	Traditional Chinese (Taiwan Mandarin)
License	Apache 2.0

Usage

With mlx-whisper

python

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="schsu/breeze-asr-25-mlx",
    language="zh",
)
print(result["text"])

With TypelessMLX App

This model is the default high-accuracy option in TypelessMLX, a macOS dictation app powered by MLX Whisper.

Performance

Optimized for Taiwan Mandarin (繁體中文 / 台灣國語)
Runs fully on-device on Apple Silicon (M1 and later)
Uses fp16 precision for faster inference on Apple Neural Engine
Significantly better accuracy on Traditional Chinese than standard Whisper Large v3

Conversion

Converted from the original PyTorch weights using weight remapping (HuggingFace Transformers → OpenAI Whisper naming convention) and tensor transposition (Conv1d channels).

Original model: MediaTek-Research/Breeze-ASR-25

Citation

If you use this model, please cite the original Breeze-ASR-25:

bibtex

@misc{breeze-asr-25,
  author = {MediaTek Research},
  title  = {Breeze-ASR-25},
  year   = {2025},
  url    = {https://huggingface.co/MediaTek-Research/Breeze-ASR-25}
}