STT · LLM · TTS — three Web Workers, one tab
Real-time Voice AI. Entirely in Your Browser.
Talk naturally. Interrupt anytime. ValaJS runs Whisper, a local LLM, and Kokoro TTS completely in your browser—no uploads, no delays, no cloud required.
Talk to it
A real conversation, right in this tab
This is the real agent, running in your browser. Click below, allow microphone access, and it'll load the actual models before saying hello.
Click below to start. First run downloads the models — after that it's near-instant.
Live — first run downloads real STT/LLM/TTS models (roughly 2GB) and caches them in your browser. Needs a WebGPU browser and mic permission. Nothing you say leaves the tab.
01
Nothing uploaded
Speech-to-text, the language model, and text-to-speech all run locally in Web Workers. No audio or text ever leaves the browser.
02
Streamed sentence by sentence
Replies are split into sentences as they generate and spoken the moment each one is ready — you hear the first line while the model is still writing the third.
03
Interruptible by design
The mic opens again before the previous turn even finishes, so barging in mid-reply stops the agent immediately, every time.
Get started
One install, three workers, zero servers
Works with Next.js, Vite, and any webpack 5 setup out of the box — no framework required.