Not a stock voice. A voice.
A Chinese-exclusive TTS voice model by Zixi Li. VITS2 lineage (forked from MeloTTS), 51.43M parameters, Chinese-only, no BERT.
- Small — 51M params is light enough for local inference, CPU included. No cloud, no API key, no GPU bill.
- Chinese-exclusive — the whole stack serves one language: 77 symbols, 6 tones, zero language-routing overhead.
- One real person's voice — a single real human's voice, so the output keeps one person's timbre, rhythm and breath.
Sample
▶ final_showcase.wav — one take, default inference parameters (sdp_ratio=0.2, noise_scale=0.6, noise_scale_w=0.8), text from Zixi Li's own fiction.
Files
config.json # model hyperparams
pytorch_model.bin # final generator weights (51.43M params, 206MB)
Usage
Code lives on GitHub: zixi-51M.
uv venv .venv --python 3.11
uv pip install git+https://github.com/lizixi-0x2F/zixi-51M
from huggingface_hub import snapshot_download
from zixi.api import TTS
ckpt_dir = snapshot_download("NoesisLab/zixi-51m")
model = TTS(language='ZH',
config_path=f'{ckpt_dir}/config.json',
ckpt_path=f'{ckpt_dir}/pytorch_model.bin')
model.tts_to_file('你好，世界。', model.hps.data.spk2id['zixi'], 'hello.wav')
# -> hello.wav, 44.1kHz Int16 mono
Model details
- Generator: 51.43M params; D/DUR trained from scratch and not needed at inference.
- Differences from upstream MeloTTS: BERT fully removed (text encoder carries only phone + tone embeddings); Chinese-only (no language embedding, no language selection; symbol table 77, tone table 6); embedding-transplant warm start (phone/tone/speaker embeddings inherited row-by-row from the MeloTTS-Chinese base by symbol name).
Data & ethics
Weights only — training data is not part of this release. The voice is the author's own; the training corpus consists of the author's own texts and is private. Please do not use this voice to impersonate the author or to put words in a real person's mouth.
Acknowledgements
Based on MeloTTS (MIT, MyShell.ai), by Wenliang Zhao, Xumin Yu and Zengyi Qin. Base weights from myshell-ai/MeloTTS-Chinese. Upstream in turn builds on coqui-ai TTS, VITS, VITS2 and Bert-VITS2.
