kotoba-whisper-v2.2 is a Japanese speech recognition model fine-tuned from OpenAI Whisper. This repository provides quantized versions optimized for CPU inference with whisper.cpp.

Available Models

Model File	Size	Quantization	Description
`kotoba-whisper-v2.2-ggml.bin`	~1.4 GB	FP16	Original precision, best quality
`kotoba-whisper-v2.2-ggml-q8_0.bin`	~780 MB	Q8_0	8-bit quantization, good balance
`kotoba-whisper-v2.2-ggml-q5_0.bin`	~513 MB	Q5_0	5-bit quantization, smaller size

Recommended Model

For best quality: Use kotoba-whisper-v2.2-ggml.bin (FP16)
For balanced performance: Use kotoba-whisper-v2.2-ggml-q8_0.bin (Q8_0)
For resource-constrained environments: Use kotoba-whisper-v2.2-ggml-q5_0.bin (Q5_0)

Usage

With whisper.cpp CLI

bash

# Clone and build whisper.cpp
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

# Download model (example with Q8_0)
# Place the downloaded .bin file in the models/ directory

# Run inference
./build/bin/whisper-cli -m models/kotoba-whisper-v2.2-ggml-q8_0.bin -f audio.wav -l ja

With Python bindings

python

from whispercpp import Whisper

w = Whisper.from_pretrained("kotoba-whisper-v2.2-ggml-q8_0.bin")
result = w.transcribe("audio.wav")
print(result)

Model Architecture

Base Model: OpenAI Whisper Large-v3
Encoder Layers: 32
Decoder Layers: 2
Hidden Size: 1280
Attention Heads: 20
Vocabulary Size: 51866
Audio Context: 1500 (30 seconds)
Mel Bins: 128

Conversion Details

These models were converted using the convert-h5-to-ggml.py script from whisper.cpp and quantized using the built-in quantize tool.

Conversion Command

bash

# Convert to GGML
python models/convert-h5-to-ggml.py <model_dir> <whisper_repo> <output_dir>

# Quantize to Q8_0
./build/bin/quantize ggml-model.bin ggml-model-q8_0.bin q8_0

# Quantize to Q5_0
./build/bin/quantize ggml-model.bin ggml-model-q5_0.bin q5_0

License

This model follows the same license as the original kotoba-whisper-v2.2 model. Please refer to kotoba-tech/kotoba-whisper-v2.2 for licensing details.

Acknowledgements

Original model: kotoba-tech/kotoba-whisper-v2.2
Inference engine: ggerganov/whisper.cpp
Base architecture: OpenAI Whisper

Citation

If you use this model, please cite the original kotoba-whisper model:

bibtex

@misc{kotoba-whisper,
  author = {Kotoba Technologies},
  title = {kotoba-whisper-v2.2},
  year = {2024},
  url = {https://huggingface.co/kotoba-tech/kotoba-whisper-v2.2}
}

0x3/kotoba-whisper-v2.2-ggml

kotoba-whisper-v2.2-ggml

Model Description