A Self-Refining Framework for Enhancing ASR Using TTS-Synthesized Data
Paper β’ 2506.11130 β’ Published β’ 5
ggml conversions of MediaTek-Research/Breeze-ASR-25 for whisper.cpp, made with whisper.cpp v1.9.4.
Original model: https://huggingface.co/MediaTek-Research/Breeze-ASR-25 β its card covers the model's languages, training and evaluation.
| File | Type | Size | Description |
|---|---|---|---|
ggml-breeze-asr-25.bin |
f16 | 3.1 GB | Source precision |
ggml-breeze-asr-25-q8_0.bin |
q8_0 | 1.7 GB | 8-bit |
ggml-breeze-asr-25-q5_0.bin |
q5_0 | 1.1 GB | 5-bit, smallest |
hf download tsuzuri-app/Breeze-ASR-25-ggml ggml-breeze-asr-25-q5_0.bin --local-dir .
whisper-cli -m ggml-breeze-asr-25-q5_0.bin -l zh -f audio.wav
Tsuzuri, a subtitle editor, offers these files as the Breeze-ASR-25 preset under Settings β Models β Transcription.
| Step | Tool |
|---|---|
| Source | breeze-asr-25.pt (fp16) at cffe7cc |
| Convert | whisper.cpp v1.9.4 convert-pt-to-ggml.py |
| Tokenizer | openai/whisper v20250625 assets |
| Quantize | whisper.cpp v1.9.4 whisper-quantize |
Every input is pinned by version and SHA256 in scripts/breeze.sh; run scripts/breeze.sh convert to rebuild the same files. The weights are unchanged apart from the format conversion and, for q8_0 and q5_0, quantization.
Apache-2.0, as the source model; see LICENSE.
@article{chou2025selfrefiningframeworkenhancingasr,
title={A Self-Refining Framework for Enhancing ASR Using TTS-Synthesized Data},
author={Cheng Kang Chou and Chan-Jan Hsu and Ho-Lam Chung and Liang-Hsuan Tseng and Hsi-Chun Cheng and Yu-Kuan Fu and Kuan Po Huang and Hung-Yi Lee},
journal={arXiv preprint arXiv:2506.11130},
year={2025}
}
Base model
openai/whisper-large-v2