levirve/Breeze-ASR-26-gguf

Breeze-ASR-26 GGUF

Unofficial GGUF conversion of MediaTek-Research/Breeze-ASR-26 for transcribe.cpp. Not affiliated with or endorsed by MediaTek Research.

Breeze-ASR-26 is a fine-tune of openai/whisper-large-v2 for Taiwanese Hokkien (Taigi / 台語). It outputs Mandarin characters, not Taigi orthography. The architecture is unchanged, so these files load as ordinary Whisper large-v2 GGUFs.

Files

File	Size
Breeze-ASR-26-F16.gguf	3.11 GB
Breeze-ASR-26-Q8_0.gguf	1.67 GB
Breeze-ASR-26-Q6_K.gguf	1.30 GB
Breeze-ASR-26-Q5_K_M.gguf	1.16 GB
Breeze-ASR-26-Q4_K_M.gguf	1.00 GB

Conversion

Source	MediaTek-Research/Breeze-ASR-26
Converter	transcribe.cpp commit 5700f127 with a local patch to scripts/convert-whisper.py
Patch	reads sharded safetensors; builds the tokenizer from vocab.json + merges.txt when tokenizer.json is absent

Convert to F32:

uv run --project scripts/envs/whisper scripts/convert-whisper.py \
  MediaTek-Research/Breeze-ASR-26 \
  --revision 7b992682e7f5ceedd0a41ebec240f01ba469d19e

Quantize:

uv run scripts/quantize-all.py models/Breeze-ASR-26/Breeze-ASR-26-F32.gguf

License

Apache-2.0, inherited from the upstream model, which is derived from OpenAI Whisper (also Apache-2.0). Full text: https://www.apache.org/licenses/LICENSE-2.0. The files in this repo are converted and quantized versions of the upstream weights; see Conversion above.