Zyphra Releases ZONOS3-Mini for Real-Time Text-to-Speech in the Browser
Technology📅 November 9, 2026👤 FreeReadText Team

Zyphra Releases ZONOS3-Mini for Real-Time Text-to-Speech in the Browser

ZONOS3-Mini can read text aloud in 43 languages on a phone or laptop. Its voice controls let a narrator change the pace or emotion of individual words within a line.

Zyphra released ZONOS3-Mini on November 9, 2026, with model weights under the Apache 2.0 license. In a browser that supports WebGPU, the first visit downloads the model. The browser then keeps a local copy and uses the device's graphics processor to generate speech.

A narrator can start with a recording of a speaker or design a voice by adjusting its age, accent and vocal texture. The model can slow down a particular word or change its emotional tone as it generates the audio. A separate editing model lets the narrator revise a sentence's delivery afterward.

The model has 800 million parameters in total. Its mixture-of-experts design uses about 100 million at each generation step. In Zyphra's browser tests, the first audio arrived in under 570 milliseconds on a Pixel 9 and under 150 milliseconds on an M3 MacBook Air.

The release also gives developers a way to serve the same voices from their own infrastructure. The NVIDIA runtime streams speech through an OpenAI-compatible endpoint, and the model weights, voice samples and browser demo are publicly available.

ZyphraZONOS3-MiniText-to-SpeechWebGPUOn-Device AINarration

来源

← Back to News