
I built a live AI avatar for desktop voice chats — where does the mouth still look wrong?
I’m building Nova, a Mac prototype that animates a face from the live audio of an AI voice conversation. The current demo uses audio from a ChatGPT/Codex voice chat; it does not read chat history or connect directly to ChatGPT or Claude model APIs yet.
The hard part has been making speech look human: closed lips staying still, teeth staying opaque during transitions, and cheeks moving with the mouth. I’m still refining lip sync and the mouth edges. I’d value specific feedback from people who work with generative characters: at what moment does the face stop feeling natural, and what would you fix first?
Demo video: https://vimeo.com/1232862019
The image is a screenshot of the current Mac prototype beside the conversation. Different avatars and voices, plus a Claude connection, are directions I want to build next; they are not shipping features today.