canto-tts-nano-v1
Open-source Hong Kong Cantonese text-to-speech, fine-tuned from MOSS-TTS-Nano (0.1B params, GPT-2 backbone, Apache-2.0). CPU-first — runs via ONNX Runtime, no GPU required.
This repo ships the ONNX export consumed by the
canto-tts Python SDK. It bundles the fine-tuned TTS
graphs, the stock (unmodified) OpenMOSS audio-codec ONNX graphs, and one baked default voice.
Why a dedicated Cantonese fine-tune
MOSS-TTS-Nano's base pretraining is Mandarin-centric — feeding it raw Hanzi produces Mandarin-accented
output. This fine-tune instead feeds jyutping phoneme tokens (onset/rime/tone) for Chinese
characters, while leaving English words as plain orthography (the base model's multilingual English
prior is already good) — this is what actually yields Cantonese output with natural English
code-switching. Text→jyutping conversion is handled upstream by
canto-hk-g2p; this checkpoint only consumes phoneme
tokens, it does not do grapheme-to-phoneme itself.
Usage
pip install canto-tts # see https://github.com/typangaa/canto-tts for current install instructions
from canto_tts import CantoTTS
tts = CantoTTS() # auto-downloads this repo via huggingface_hub, cached under ~/.cache/huggingface
tts.synthesize("多謝晒，今日天氣幾好。", "hello.wav")
# English code-switching
tts.synthesize("我哋一齊去 IFC food court 食飯。", "codeswitching.wav")
Advanced / no-SDK usage: browser_poc_manifest.json describes the ONNX graph layout (prefill/decode
steps for the TTS local+global transformer, plus the codec encode/decode graphs) and carries the
baked default-voice audio_codes — see canto_tts.backends.onnx_backend.OnnxBackend in the SDK repo
for the reference ONNX Runtime session-wiring implementation.
Gate results (N=5 repeats, Qwen3-ASR-1.7B judge)
| Gate | Set | Result | Threshold | Status |
|---|---|---|---|---|
| CER | golden_prompts (100 Cantonese sentences) | 11.82% | < 15% | ✅ |
| Tone accuracy | same | 84.22% | ≥ 70% | ✅ |
| Code-switch CER | same | 13.87% | < 15% | ✅ |
| Pure-English CER | 20 English-only sentences | 9.7% | < 15% | ✅ |
| Voice-clone CER (zero-shot) | 6 speakers × 5 targets | 10.2% | ≤ 11% | ✅ |
Full lineage, methodology, and paired-bootstrap significance notes: see the GitHub repo.
Default voice
The single baked default voice is this checkpoint's own unconditional (mode="continuation")
output — i.e. self-generated speech, not a copy of any identifiable real person's recording and not
borrowed from another project's demo asset. This sidesteps voice-likeness/rights questions entirely
for the shipped default. There is currently no runtime voice selection in the ONNX bundle (one baked
voice only) — multi-voice picking is planned for a future release via the underlying checkpoint's
native zero-shot voice-cloning path, not this static ONNX export.
Known limitations
- Single default voice — no voice selection or cloning in this ONNX bundle.
- Cantonese + English code-switching only — Mandarin capability was intentionally not preserved (the phoneme fine-tune overrides the base model's Mandarin reading prior for Chinese characters by design).
- No explicit style/emotion/gender control at inference — the training schema's control fields are fixed to a neutral default for this checkpoint.
License
Apache-2.0, inherited from MOSS-TTS-Nano (OpenMOSS). The
audio codec graphs (MOSS-Audio-Tokenizer-Nano-ONNX/) are the stock, unmodified OpenMOSS codec,
included here only for convenience so the SDK has a single download.
Training data is privately sourced and not released (copyright) — only these weights are public.