Introducing LLaMA Voice Chat

https://twitter.com/ggerganov/status/1640022482307502085

by iKlsR • 3 years ago

85 5 3 years ago

mdaniel• 3 years ago

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

fshbbdssbbgdd• 3 years ago

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.

happybuy• 3 years ago

The conversation back and forth with a bit of AI whimsy reminds me a lot of Apple's Knowledge Navigator video: https://www.youtube.com/watch?v=umJsITGzXd0

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.

mrverify• 3 years ago

Anyone build this? I'm not familiar with the cmake framework.

jarbus• 3 years ago

I love it, and gg’s work, but damn that latency is painful

Introducing LLaMA Voice Chat

Related Stories