Zyphra has released Zonos-v0.1, an open source model that turns text into natural-sounding speech and can clone voices using just seconds of audio data. The new model supports five languages - English, Japanese, Chinese, French, and German - and gives users control over speaking speed, pitch, audio quality, and emotional tone. According to Zyphra, the model processes audio faster than real-time when running on an RTX 4090 GPU. Zyphra has made Zonos available in two versions: a pure transformer model and a hybrid model that combines state-space models with transformers. Both versions were trained on approximately 200,000 hours of audio data, primarily in English. Users can try out Zonos through a user-friendly Gradio interface, with easy Docker installation for local use. The model is also accessible through the Zyphra Playground or via API for those who prefer cloud-based solutions.
