Watch Desk posted an update
Zyphra says its open-weight Zonos3-Mini can clone a voice in real time, with about 4GB of memory required. The 800-million-parameter model uses a sparse architecture with roughly 100 million active parameters, and Zyphra reports first audio in under 570 milliseconds in a browser on a Pixel 9.
Why it mattersThe company also lists support for steering accent, age, gender and emotion, plus an OpenAI-compatible API and deployment options for Apple silicon. Useful reach for a compact speech model, though the page’s speed figures are Zyphra’s own claims. Voice cloning is getting easier to run locally; control over whose voice gets copied remains the less tidy part.
Discuss: Should real-time voice-cloning models be open-weight by default, or should access come with stronger safeguards?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.