canto-tts

Cantonese Hong Kong Text-to-Speech · CPU-first · Apache-2.0

An open-source Cantonese (Hong Kong) text-to-speech SDK. Runs on CPU via ONNX Runtime — no GPU required. Fine-tuned from MOSS-TTS-Nano (0.1 B params, GPT-2 backbone, Apache-2.0, by OpenMOSS).

Published — CER 11.82% / tone accuracy 84.22% / code-switch CER 13.87% on a 100-sentence gate set (N=5 repeat eval). Weights on HuggingFace. Single default voice only — no voice cloning, no voice selection (yet).

Install

pip install canto-tts

Quickstart

from canto_tts import CantoTTS

tts = CantoTTS()  # auto-downloads typangaa/canto-tts-nano from HuggingFace on first use
tts.synthesize("多謝晒，今日天氣幾好。", "hello.wav")
print("Saved to hello.wav")

# English code-switching works too
tts.synthesize("我哋一齊去 IFC food court 食飯。", "codeswitching.wav")

Weights auto-download from HuggingFace and cache via the Hub (~/.cache/huggingface/hub/).

Quality Modes (opt-in)

Two opt-in quality= modes trade extra compute for a more reliable draw, without touching the model:

tts.synthesize(text, "out.wav", quality="duration_filter", max_attempts=3)   # no extra deps
tts.synthesize(text, "out.wav", quality="best_of_n", best_of_n=4)            # needs canto-tts[quality]

quality="best_of_n" reranks N draws by ASR character-error-rate; pass asr_backend="sensevoice" (needs canto-tts[quality-sensevoice]) for ~7x faster reranking at a small accuracy cost. Full writeup: GitHub README.

Limitations

Model & License

• Base model: MOSS-TTS-Nano by OpenMOSS
• Weights: published at huggingface.co/typangaa/canto-tts-nano
• License: Apache-2.0 (inherited from MOSS-TTS-Nano)

Full Documentation

Full README, CLI docs, Docker self-hosting, and deployment guide: https://github.com/typangaa/canto-tts