Hacker News Reader

But AI is currently failing at American Sign Language, notably at facial expressions, miserably.

Yes, we are working on that but I see writing on the wall and it is not soon enough.

ASL seems like a really hard problem. I've learned a fair bit myself being friends with a few interpreters and deaf people and two different people can sign the same thing in a way that would look very different to an AI. Sometimes it feels like you have to put a fair bit of effort into understanding how a specific person signs (from my very inexperienced perspective). I'm curious if this could be overcome with sufficient data, but where is this massive archive of videos of sign? There are some databases certainly, but nothing close to the level of which we have written and spoken english via the internet. Plus then you get into regional and cultural dialects... i think banks will be obligated to hire interpreters for the forseeable future.

Gallaudet University has the video archives.

Here is one of many:

https://media.gallaudet.edu/channel/Videolibrary+Archive/158...

I'm very aware of these archives. They are very clearly not sufficient to train a model to interpret asl to the standards of corporate america. Let alone the standards of an arbitrary human. Hence my above post.

Reply on news.ycombinator.com