I've worked on a project that was burned by this, everything works in the lab, then we start taking it home and nothing works with kids.
Standard speech reco and face reco either aren't optimized for children, or don't have children in their training data at all. Particularly tricky for speech if there is a grammar model...
Technically this sounds really impressive, but I'm a bit confused by that choice of demo. I'm not sure if I'm missing something but I basically got a question posed by the computer with no context - so it could just be precanned, and the rest of the time was just the kid talking.
If the system can reply to that in a smart and in character way then that's incredible but that demo just doesn't show anything they talked about.
On the other hand their character voicing sessions are the best example of do things that don't scale I've heard of. :)
I interviewed at ToyTalk when they were first starting out. Oren and Martin are both incredibly smart (both for what they build and for not offering me a job).
I love the combination of storytelling and technology that they've been building.
Standard speech reco and face reco either aren't optimized for children, or don't have children in their training data at all. Particularly tricky for speech if there is a grammar model...