Supertonic-ZH
An unofficial Mandarin Chinese fine-tune of the Supertonic-3 on-device text-to-speech model, distributed as optimized ONNX graphs with streaming synthesis and Chinese-specific text normalization for dates, money and phone numbers.
Supertonic-ZH extends the Supertonic-3 text-to-speech family, which does not synthesize Mandarin out of the box, with a Mandarin fine-tune that runs as fused, optimized ONNX graphs in pure ONNX Runtime, while reusing the frozen, unmodified upstream vocoder. It handles tones, polyphones, erhua and neutral-tone and yi-bu sandhi in a consistent preset voice, and normalizes dates, money, percentages, phone numbers and full-width punctuation to how they would be spoken.
The authors report roughly 10x real-time generation on GPU and describe it as lightweight and built for on-device, offline, low-latency use, with streaming synthesis for low first-audio latency and barge-in support in real-time dialogue. A public 100-sentence Chinese test set and its character error rate results are included for reproducibility, alongside example audio in Mandarin and several other languages carried over from the base model.