The details
The system chains separate models: Nemotron speech recognition turns speech into text, the small Bolt E4B model writes replies, Magpie voices them from short recorded samples, and Audio2Face-3D moves the mouth. After testing more than a dozen models, the author chose Bolt partly for its smaller footprint. Prepared greetings and saved audio handle common encounters. Real-world use is still unproven: costume recognition has been shown only on a synthetic video of sheep, and full performance testing still needs real visitors. The project’s own code is not packaged for release.
Why it matters
This is a vendor demo, written by an employee who had early access, so its numbers are not independent. Still, it shows one direction for voice agents: modest models on local hardware, with personality added through custom training and ordinary software tracking the conversation. Its fastest replies depend on reusing prepared lines, which suggests live generation alone may still feel slow.