Zonos-v0.1 is an open-weight text-to-speech model designed for highly natural and expressive speech generation, including zero-shot voice cloning. It targets researchers and developers seeking high-quality, controllable TTS capabilities, offering performance comparable to commercial providers.

How It Works

Zonos utilizes a transformer or hybrid backbone for DAC token prediction, preceded by text normalization and phonemization via eSpeak. This architecture allows for conditioning on speaker embeddings or audio prefixes, enabling fine-grained control over speech rate, pitch, audio quality, and emotions. The model outputs audio natively at 44kHz.

Quick Start & Requirements

• Install: uv sync (or uv sync --extra compile for hybrid) followed by uv pip install -e . (or .[compile]).
• Prerequisites: Linux (Ubuntu 22.04/24.04 recommended), macOS. GPU with 6GB+ VRAM (Nvidia 3000-series+ for hybrid). eSpeak-ng library.
• Resources: CPU-only is possible but slow. Docker installation is available.
• Demo: playground.zyphra.com/audio

Highlighted Details

• Zero-shot TTS with voice cloning from short audio samples.
• Supports multilingual generation (English, Japanese, Chinese, French, German).
• Fine-grained control over speaking rate, pitch, audio quality, and emotions (happiness, fear, sadness, anger).
• Real-time factor of ~2x on an RTX 4090.
• Includes a Gradio WebUI for easy speech generation.

Maintenance & Community

• No specific contributors, sponsorships, or community links (Discord/Slack, roadmap) are mentioned in the README.

Licensing & Compatibility

• The README does not explicitly state a license. Compatibility for commercial or closed-source use is not specified.

Limitations & Caveats

• Experimental Windows support is available via a fork. The hybrid model has specific GPU requirements. The README does not detail any known bugs or deprecations.