Kitten TTS Native

Native C++ inference engine for Kitten TTS + Python GGUF model conversion tools.

Project Structure

├── cpp/           C++ inference engine (ggml backend, CMake build)
│   ├── src/        kitten-tts.cpp (library) + main.cpp (CLI)
│   ├── include/    kitten-tts.h (public API)
│   └── extern/     ggml + espeak-ng (git submodules)
├── python/        GGUF conversion scripts
│   ├── convert_gguf.py   ONNX → GGUF converter
│   ├── convert_onnx.py   ONNX weight extraction
│   └── model.py          PyTorch model definition
└── requirements-gguf.txt  Python dependencies for GGUF conversion

Build C++ CLI

# Clone with submodules
git clone --recursive https://github.com/YOUR_USER/kitten-tts-native.git
# or if already cloned:
git submodule update --init --recursive

# Build
cd cpp
cmake -B build
cmake --build build -j$(nproc)

# Run
./build/kitten-tts-cli --model ../path/to/model.gguf --text "Hello world" --output hello.wav

CLI Options

--model PATH      GGUF model path (required)
--text STRING     Text to synthesize (required, or --input-ids)
--input-ids LIST  Comma-separated token IDs (alternative to --text)
--voice NAME      Voice: Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, Leo (default: Bella)
--ref-id INT      Reference style ID 0-399 (default: 0)
--speed FLOAT     Speed 0.5-2.0 (default: 1.0)
--output PATH     Output WAV path (default: output.wav)
--threads INT     Thread count (default: 4)

Convert Model to GGUF

pip install -r python/requirements-gguf.txt

python python/convert_gguf.py \
    --onnx path/to/model.onnx \
    --voices path/to/voices.npz \
    --output model.gguf

Dependencies

• C++: CMake 3.14+, C++17 compiler, ggml (submodule), espeak-ng (submodule)
• Python: PyTorch 2.0+, NumPy, ONNX

License

Apache 2.0