GitHub
One human voice, 51M params, runs on your CPU
Start new thread

GitHub - One human voice, 51M params, runs on your CPU

by•
Open-weights Chinese TTS: 51.43M params, VITS2, no BERT. Runs on a laptop CPU — no cloud, no API key. One real person's voice, not a stockvoice or a multi-speaker average. MIT. Weights on Hugging Face, code on GitHub.

Add a comment

Replies

Best
Maker
📌
Maker here. I'm a writer, and the voice in this model is my own — it was trained on my own texts, which is why the weights are open (MIT)but the training data stays private. I wanted a TTS I could actually own: small enough to run on a laptop CPU, no account, no API key, no meter running. The tradeoffs aredeliberate — Chinese-only and single-speaker, so the whole stack serves exactly one language and one voice. No BERT, no language routing, 77symbols, 6 tones. What you get: the final generator weights (51.43M params, 206MB) and the full training/inference code. What you don't get: my corpus, andone request that comes with the voice — please don't use it to put words in a real person's mouth. Weights: Code: