Siri could've been this but overtime it has become more and more neutered to the point that it feels like talking to a drunken command line interface.
It seems like ASR is interpolating the words since at 1:04 it sure does sound to me like Georgi says "yeah, not sure how to feel about this" but it is transcribed as "Yeah, I'm not sure how to feel about this". It also seems to be doing "um" and "uh" suppression, which I can see how that's valuable in some circumstances but "do what I meant" rarely ends well in my experience
ASR systems usually have a language model. It improves accuracy a lot because there are a lot of similar-sounding words hard to tell apart you can eliminate because of context. Depending on the performance, you may have to lean heavily on the language model to get decent results. Even though it sometimes causes errors, it prevents more.
I love it, and gg’s work, but damn that latency is painful
Anyone build this? I'm not familiar with the cmake framework.