Any2Canto 🎙️

Any2Canto turns text in any language into natural spoken Cantonese using OmniVoice zero-shot voice cloning. Optional LLM translation produces colloquial Cantonese before synthesis; Chinese input is normalized with OpenCC (Simplified → Traditional, Hong Kong glyphs).

✨ Features

• 🌐 Multilingual input — Type in English, Mandarin, Japanese, Korean, etc.; non-Chinese text can be translated to colloquial Cantonese via an LLM API (when configured).
• 🎭 Voice cloning — Upload 3–10 seconds of reference speech.
• 🧭 Cantonese-first pipeline — OpenCC s2hk + language detection to avoid unnecessary API calls for written Chinese.

📂 Repository layout

🚀 Quick start (local)

From the space/ directory:

cd space
pip install -r requirements.txt

Set environment variables as needed:

Run the UI:

python app.py

💡 Local run: python app.py starts Gradio on localhost (default URL in the terminal). Device is picked automatically (cuda → mps → cpu); on Mac with Apple Silicon, install PyTorch for macOS (do not use the CUDA wheel index in requirements.txt—install torch / torchaudio from pytorch.org or pip install torch torchaudio on Mac). On Hugging Face Spaces, spaces provides the GPU decorator; locally the app runs without that package.

🤗 Hugging Face Space

The file space/README.md contains the Space configuration front matter (app_file, sdk, etc.). Point your Space repository at space/ as the app root, or sync this subtree to your Space repo.

🙏 Credits

• OmniVoice — k2-fsa/OmniVoice

⚖️ License

This repository follows the same license as OmniVoice: Apache License 2.0. The license text is the upstream OmniVoice LICENSE. Redistribution must include that file and NOTICE (third-party and attribution).