A new 2B open-source TTS model clones voices from seconds of audio

A new lightweight text-to-speech model just landed in the open-source scene, packing impressive zero-shot voice cloning capabilities into a compact 2-billion-parameter architecture. Feed it just a few seconds of raw audio, and it will read out any text in that exact voice.

Capturing whispers and stutters on consumer GPUs

What makes this model truly stand out is its knack for expressive nuance. While many speech models produce overly sterile outputs, this one can faithfully preserve subtle quirks like whispers and nervous stutters.

It also handles cross-lingual synthesis with ease. Even if your reference audio only speaks one language, the model can generate convincing speech in other languages using that same vocal identity. In benchmark evaluations plotting error rates against speaker similarity, it landed in the ideal top-left sweet spot, outperforming recent rivals like CTDTS, Vox CPM, Qwen 3 TTS, and Index TTS 2.

The weights and code are available on GitHub under an Apache 2.0 license, meaning anyone can use it commercially. At around 5GB, the base model easily fits inside consumer-grade graphics cards for fully local inference.

Why it matters

High-fidelity voice synthesis is no longer the exclusive playground of expensive cloud APIs. For independent creators and developers, having a sub-5GB model that captures authentic human delivery locally—without per-token fees or privacy compromises—makes international dubbing and synthetic voice generation trivial.

We think releasing this under Apache 2.0 puts serious pressure on closed-source voice generation providers. But there is a flip side: as high-end cloning from seconds of sample audio becomes runnable on home hardware, audio verification and anti-spoofing just got a lot harder.