Zyphra ZONOS3-Mini: Real-Time Voice Cloning on Your Phone with This Open-Source TTS Model

AI startup Zyphra has released ZONOS3-Mini, an open-source text-to-speech (TTS) model that enables real-time voice cloning directly on mobile phones and within web browsers. This compact model, announced on November 9-10, 2026, supports 43 languages for on-device real-time cloning and offers high-fidelity speech generation with customizable voice characteristics.

Introducing ZONOS3-Mini: On-Device Voice Synthesis

ZONOS3-Mini stands out as a compact, sparse mixture-of-experts model. It features approximately 100 million active parameters from a total of 800 million, making it efficient for local execution. The model was developed using 112 NVIDIA H100 GPUs, completing its training in about 2.5 days. This architecture enables the model to generate natural speech and perform high-fidelity voice cloning from a single audio sample.

The model's capabilities extend to creating custom character voices, allowing users to adjust attributes such as age, accent, and vocal texture. Furthermore, ZONOS3-Mini provides word-level control over elements like emotion, pace, and delivery, offering granular customization for synthesized speech.

Real-Time Performance and Accessibility

A key feature of ZONOS3-Mini is its ability to operate in real time on various devices, including phones and laptops. This includes direct execution within web browsers through WebGPU, enhancing its accessibility for developers and users. The model's weights are publicly available on Hugging Face under the Apache 2.0 license, promoting its integration into diverse applications and projects.

The model supports real-time voice cloning across 43 languages, making it a versatile tool for global applications. Its on-device processing capability reduces reliance on cloud infrastructure, potentially offering benefits in terms of privacy and latency for audio generation tasks.

Competitive Landscape and Technical Details

Zyphra has published comparisons of ZONOS3-Mini against other models, including Fish Audio, Qwen, Cartesia, ElevenLabs, and OmniVoice. These comparisons focused on metrics such as latency and throughput, indicating Zyphra's positioning of ZONOS3-Mini within the competitive landscape of text-to-speech and voice cloning technologies.

Technically, ZONOS3-Mini is an 11-layer language model that incorporates gated-DeltaNet layers with global attention. It processes SuperBPE text tokens and predicts 10 × 4096 dual-codec audio codes at 15 Hz. A causal multi-resolution iSTFT decoder then converts these codes into 44.1 kHz audio, ensuring high-quality output.

Key Takeaways for Developers and Users

The release of ZONOS3-Mini provides an open-source option for developers seeking to integrate advanced voice synthesis and cloning into their applications. Its on-device, real-time capabilities, combined with extensive language support and granular control over speech characteristics, make it a notable development in the field of audio AI.

The availability of its weights under an open-source license on Hugging Face encourages community contributions and broader adoption. This model could be particularly relevant for applications requiring localized speech generation or personalized voice interfaces without constant internet connectivity.

Conclusion

Zyphra's ZONOS3-Mini represents a significant step in making advanced voice cloning and text-to-speech technology more accessible and efficient for on-device applications. Its open-source nature and real-time performance across multiple languages position it as a valuable tool for developers and researchers exploring the frontiers of audio AI.

Sources

• Zyphra/ZONOS3-Mini · Hugging Face
• sglang-omni/docs/cookbook/zonos3.md at main
• GitHub - Zyphra/ZONOS3-Mini · GitHub
• mlx-community/ZONOS3-Mini · Hugging Face
• ZONOS3-Mini | Zyphra