Zyphra has released ZONOS2 — the second generation of its text-to-speech (TTS) model, which utilizes a Mixture of Experts (MoE) architecture to achieve a balance between studio-quality sound and minimal generation latency.
What Happened
The ZONOS2 model has 8 billion parameters, of which 900 million are active during inference. It provides audio generation at a 44.1 kHz sampling rate via the Descript Audio Codec (DAC) in real-time. One of the key technical features is the use of byte tokenization (raw UTF-8 bytes), which allows the model to support multiple languages without the need for complex phonetic preprocessing and effectively handle code-switching.
Context
In the field of speech synthesis, developers traditionally face a trade-off between voice cloning quality and speed (latency). The use of MoE architecture and the transition to end-to-end byte models allow for a move away from classic multi-stage pipelines in favor of more flexible and faster systems.
Why It Matters for the Industry
The release of ZONOS2 in the open-source segment significantly lowers the barrier to entry for creating advanced voice products, freeing developers from dependence on proprietary APIs and massive cloud budgets. The MoE architecture allows for scaling synthesis capabilities without a proportional increase in computational costs per request, paving the way for the mass adoption of high-quality voice interfaces.
Why It Matters for Users
Users and developers gain access to a tool for creating highly natural digital voices with support for Russian, English, and other languages at studio quality. The model supports "stable" and "expressive" modes, allowing users to choose between signal purity and emotional expressiveness, which is critical for creating lifelike AI agents.
What Is Not Yet Known / Limitations
For full production use, additional data is required regarding stability under load, precise inference costs, and actual latency metrics across various scenarios.
Sources
Author
Look at AI, Editorial Team