Breeze-ASR-26-GGML

This repo contains GGML exports of MediaTek-Research/Breeze-ASR-26 for use with whisper.cpp.

The base model is a Whisper large-v2 fine-tune for Taiwanese Hokkien / Taigi ASR, producing mostly Mandarin Chinese characters.

Included files

File	Format	Size	SHA256
ggml-model.bin	fp16 GGML	2.9 GB	6d58f81d79155deb5037f995a048856f6deaa9e06f59a89183cc421fa37cb1ad
ggml-model-q5_0.bin	q5_0 GGML	1.0 GB	60f25e3a21feca12ec082e6d36f08f94455d9900d6343f7fcb2906f71cc7c449
ggml-model-q8_0.bin	q8_0 GGML	1.5 GB	e878f611bf3ef407cff3c1e2ecf5f60dc92bc6dd09a85b22166731c0d92764a1

Checksums are also provided in SHA256SUMS.

Usage with whisper.cpp

Build whisper.cpp, then run:

./build/bin/whisper-cli \
  -m ggml-model-q5_0.bin \
  -f audio.wav \
  -l zh

Or use the fp16 model:

./build/bin/whisper-cli \
  -m ggml-model.bin \
  -f audio.wav \
  -l zh

Notes

• whisper.cpp currently uses GGML model files for Whisper models.
• I am not publishing a GGUF variant here because that is not the standard/runtime-compatible Whisper artifact for whisper.cpp today.
• If an official Whisper GGUF path becomes stable, this export pipeline can be extended.

Provenance

• Source model: MediaTek-Research/Breeze-ASR-26
• Conversion path: HF Transformers checkpoint → whisper.cpp GGML
• Quantization: whisper-quantize with q5_0 and q8_0