fredchu/breeze-asr-26-mlx-fp16

Breeze-ASR-26 — MLX (fp16)

Apple MLX port of MediaTek-Research/Breeze-ASR-26, the Taiwanese Hokkien (Taigi) ASR model from MediaTek Research's MR Breeze 3 series.

Runs on Apple Silicon Macs (M1/M2/M3/M4) via the mlx-whisper package — same API as mlx-community/whisper-* checkpoints.

Files

file	size
weights.safetensors	~3.1 GB
config.json	<1 KB

Usage

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="fredchu/breeze-asr-26-mlx-fp16",
    language="zh",
)
print(result["text"])

CLI:

mlx_whisper audio.wav --model fredchu/breeze-asr-26-mlx-fp16 --language zh

Performance (M1 Max, mlx-whisper)

sample	duration	RTF
Mandarin financial speech (60s)	60.0s	8.72× real-time
Taiwanese Hokkien sample (25s)	25.0s	6.28× real-time

When to use this vs. the 4-bit variant

A companion fredchu/breeze-asr-26-mlx-4bit (877 MB, palette quantized) is also available.

In our 2026-04-30 evaluation on a real Taigi sample (a Mandarin-speaking creator using the Taigi word "漂泊" pio-pôa), the 4-bit variant transcribed "漂泊" correctly while the fp16 variant produced "瀟灑" instead — a counterintuitive result, possibly because palette quantization re-calibrates outlier weights in a way that helps generalize to underrepresented Taigi tokens. Worth verifying if you have a Taigi-heavy use case.

For pure Mandarin or read benchmarks, fp16 should remain the safer choice.

Limitations (inherited from base model)

• Outputs Mandarin Chinese characters, not Taigi orthography (台語正字 / 台羅)
• Trained on ~10,000 hours of synthetic Taigi speech — distribution gap with real spontaneous speech
• English brand/proper nouns are aggressively transliterated: in our Mandarin test, Hello became 哈囉, Austin became Alsted, Netflix became Nathalie 的時事. ASR-25 (MediaTek-Research/Breeze-ASR-25) handles these correctly. Do not use this model for content with frequent English code-switching.
• All segments come back as one ~30-second block regardless of audio content (model training behaviour, not framework setting). Post-process if you need finer subtitle granularity.

Conversion

Built with a custom wrapper around mlx-examples/whisper/convert.py that adds sharded-safetensors loader support (the source repo ships weights as 5 GB + 1 GB shards, which the upstream converter doesn't handle).

The conversion script is open source on GitHub — search for convert_breeze_asr_26.py in For_Claude/scripts/asr-eval/.

License

Apache 2.0 — inherits from the base model.

Acknowledgments

• MediaTek Research for the original Breeze-ASR-26 weights
• Apple MLX team for the framework and conversion tooling