Zyphra Releases Open-Source Real-Time TTS 'ZONOS2'

Zyphra has unveiled ZONOS2, its next-generation real-time text-to-speech (TTS) model. This model is an open-source model provided under the Apache 2.0 License, designed to solve the trade-off between quality and speed.

Key Technical Features:

• Sparse MoE (Mixture of Experts) Architecture: While it has a total of 8B parameters, it maximizes efficiency by using only 900M active parameters during inference.
• High-Fidelity Audio Generation: It generates 44.1 kHz studio-quality audio by predicting Descript Audio Codec (DAC) tokens.
• Language Support and Code-Switching: By directly reading raw UTF-8 bytes without a separate phonemizer, it enhances performance for low-resource languages such as Korean, Chinese, and Japanese, and supports mid-sentence code-switching.
• Zero-shot Cloning: It offers powerful voice cloning performance that captures a speaker's unique characteristics without additional fine-tuning.

ZONOS2 was trained on a massive audio dataset of over 6 million hours, and it demonstrated high expressiveness by recording superior Prosody scores compared to existing major TTS models (Qwen 3 TTS, Cartesia Sonic 3.5, etc.).