illitan/whisper-large-v3-turbo-zh-TW-ggml

Whisper Large v3 Turbo zh-TW (GGML)

GGML format models for whisper.cpp, converted from JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW.

Model Information

• Base Model: OpenAI Whisper Large v3 Turbo (0.8B parameters)
• Fine-tuned on: Common Voice 19.0 zh-TW (44 hours of audio)
• Language: Traditional Chinese (Taiwan)
• CER: 8.6% (Character Error Rate)
• License: MIT

Available Files

File	Size	Format	Description
ggml-whisper-large-v3-turbo-zh-TW.bin	1.6GB	F16	Full precision, best quality
ggml-whisper-large-v3-turbo-zh-TW-q8_0.bin	874MB	Q8_0	Quantized, recommended for mobile devices

Usage with whisper.cpp

# Clone whisper.cpp
git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp

# Build
cmake -B build && cmake --build build --config Release

# Download model
wget https://huggingface.co/leaker/whisper-large-v3-turbo-zh-TW-ggml/resolve/main/ggml-whisper-large-v3-turbo-zh-TW-q8_0.bin

# Run inference
./build/bin/whisper-cli -m ggml-whisper-large-v3-turbo-zh-TW-q8_0.bin -f audio.wav -l zh

Conversion

These models were converted using the convert-h5-to-ggml.py script from whisper.cpp:

python3 ./models/convert-h5-to-ggml.py \
  /path/to/JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW \
  /path/to/openai/whisper \
  /output/path

# Quantize to Q8_0
./build/bin/whisper-quantize input.bin output-q8_0.bin q8_0

Credits

• Original model: JacobLinCool/whisper-large-v3-turbo-common_voice_19_0-zh-TW
• whisper.cpp: ggml-org/whisper.cpp