A pure Swift implementation for Kokoro TTS (Text-to-Speech) using local ONNX models, supporting multiple quantization formats and automatic cache management.
This project is currently under active development and features are not yet fully stable.
Tokenization quality: The current tokenization implementation has limited accuracy for long sentences and complex words. Chinese language support is not yet implemented.
This project is different from kokoro-ios:
- ONNX model format and runtime: Uses ONNX Runtime for model inference, while kokoro-ios uses MLX Swift runtime.
MIT
- Kokoro-82M - Original model
- ONNX Community - ONNX model conversion
Issues and Pull Requests are welcome!