Voice agents with Realtime Video — Sidney Primas, LemonSlice
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
A conversational video avatar must generate its next movement before the rest of the conversation exists. LemonSlice, which builds interactive avatars, addresses this by training video generation to use only preceding frames. The published description of its technical talk explains the problem: errors in one generated block can carry into later blocks and accumulate during a long session.
The description also reports compressing roughly thirty denoising steps into one to support real-time generation. Together, these choices address continuity and response speed, two requirements for keeping an avatar in a live conversation. They explain the engineering approach without establishing a measured thirtyfold speedup.
- attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
Checking sign-in…
Loading comments…






