Brief

Zyphra has released ZONOS2, a new text-to-speech model trained on over 6 million hours of multilingual speech data. The model aims to match or exceed the quality and expressiveness of leading TTS providers while maintaining low latency. ZONOS2 is particularly adept at high-fidelity voice cloning and utilizes a Mixture-of-Experts (MoE) architecture for its backbone. AI

Zyphra has released ZONOS2, an open-source, real-time text-to-speech model featuring 8 billion total parameters and 900 million active parameters for efficient inference. This sparse Mixture-of-Experts model excels at high-fidelity, zero-shot voice cloning and aims to overcome the typical trade-off between speech quality and speed. ZONOS2 processes raw UTF-8 bytes instead of phonemes, improving support for multiple languages and code-switching, and was trained on over 6 million hours of audio data. AI