Zonos 2 is an open-weights text-to-speech model from Zyphra, trained on more than 6 million hours of multilingual speech, with a Mixture-of-Experts backbone and low latency. It clones a voice from a few seconds of reference audio and speaks the given text in that voice.