Breeze ASR 25 is an advanced ASR model fine-tuned from Whisper-large-v2, optimized for Taiwanese Mandarin and Mandarin-English code-switching scenarios.

Key Features

Traditional Chinese direct output (no OpenCC post-processing needed)
Mandarin-English code-switching: 55.9% WER reduction vs Whisper-large-v2
Taiwanese Mandarin optimized: CommonVoice16-zh-TW WER 7.97 (vs 9.84 baseline)
MLX native: runs on Apple Silicon GPU via Metal

Use with mlx-whisper

pip install mlx-whisper

python

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="Kenji8000/Breeze-ASR-25-mlx",
    language="zh",
)
print(result["text"])

Performance

Tested on Mac Studio M2 Ultra 128GB, mlx-whisper 0.4.3, mlx 0.31.1:

Metric	Value
RTF (Real-Time Factor)	0.08x
Inference (model cached)	0.52s for 6.3s audio
Memory (RSS)	~3.1 GB

Conversion

Converted using mlx-examples/whisper/convert.py with dtype float16.

bash

python convert.py --torch-name-or-path MediaTek-Research/Breeze-ASR-25 \
    --mlx-path mlx_models/breeze-asr-25 --dtype float16
# Note: rename output model.safetensors → weights.safetensors for mlx-whisper

Original Model

Base: openai/whisper-large-v2
Fine-tuned by: MediaTek Research
License: Apache 2.0
Paper: arXiv:2506.11130