Automatic Speech Recognition
MLX
whisper
mlx-whisper
taiwanese-hokkien
taigi
zh
nan
8-bit precision
quantized
Instructions to use doggy8088/Breeze-ASR-26-MLX-8bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use doggy8088/Breeze-ASR-26-MLX-8bit with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir Breeze-ASR-26-MLX-8bit doggy8088/Breeze-ASR-26-MLX-8bit
- Notebooks
- Google Colab
- Kaggle
- Local Apps
- LM Studio
Breeze-ASR-26-MLX-8bit
這是 doggy8088/Breeze-ASR-26-MLX 的 8-bit 量化版, 由 MLX fp16 checkpoint 再量化而成,適合在 Apple Silicon 上降低記憶體占用。
原始來源模型為 MediaTek-Research/Breeze-ASR-26。 這是一個以 Whisper Large-v2 為基礎微調的台語(Taigi / Taiwanese Hokkien)ASR 模型。
量化設定
- bits: 8
- group size: 64
- quantization mode: MLX affine weight-only quantization
使用方式
pip install -U mlx-whisper
CLI:
mlx_whisper audio.wav --model doggy8088/Breeze-ASR-26-MLX-8bit --language zh
Python:
import mlx_whisper
result = mlx_whisper.transcribe(
"audio.wav",
path_or_hf_repo="doggy8088/Breeze-ASR-26-MLX-8bit",
language="zh",
)
print(result["text"])
注意事項
- 這個模型主要用於台語 ASR,但輸出偏向華語中文字。
- 建議使用
language="zh",或省略讓模型自動偵測。 - 若你想要最高精度,請改用 fp16 原版:doggy8088/Breeze-ASR-26-MLX
- Downloads last month
- 20
Hardware compatibility
Log In to add your hardware
8-bit
Model tree for doggy8088/Breeze-ASR-26-MLX-8bit
Base model
openai/whisper-large-v2 Finetuned
MediaTek-Research/Breeze-ASR-26 Finetuned
doggy8088/Breeze-ASR-26-MLX