ZeroWeight LogoZeroWeight AI
Login
ZeroWeight AI

Blog

Model releases, engineering notes and product updates from the team building real-time AI characters, and the open speech models underneath them.

Model release19 August 2026 · 9 min read

ZeroTTS: Vietnamese speech that starts talking in 70 milliseconds

An open 202M-parameter zero-shot text-to-speech model for Vietnamese, built for real-time voice agents. It streams from a laptop CPU without trading away the prosody that lightweight TTS usually gives up.

1.03%
word error rate
70 ms
to first audio
202M
parameters
Read it
Open source on GitHubThe models and runtimes we write about, in full.Walkthroughs on YouTubeBuilds and tutorials, if you would rather watch it happen.