Okay, this is a bit immature and you’ll learn exactly nothing from this, but I could not resist. I’m following the “Kubernetes Course from a DevOps guru” course on Udemy.com. The videos on Udemy are simply Youtube videos. Just like regular Youtube, you can turn on close captions.
If the publisher of a Youtube video did not enter typed captions, Youtube will use machine learning algorithms to automatically create them instead. This works well when people speak fluent and accentless English (and maybe other languages as well). But results get a bit off even if someone has only a slight Scottish accent, like one of my Youtube favorites and Kerbal Space Program pilot, Scott Manley.
The guys narrating this Kubernetes course speak fairly accentless English and articulate well, I think. But when they say “Kubernetes” (invented at Google, mind you), it’s clear that the closed caption algorithm didn’t have that word in the training data. And I keep snickering about what it instead comes up with.
So here are the things Youtube’s closed caption algorithm “hears” when “Kubernetes” is pronounced:
- Kubert at ease
- Coober Nettie’s
- Coober nighties
- Kluber Nettie’s
- Cuban Nettie’s
- Kuber Nettie’s
- Cuber Noddy’s
- Abernathy’s
- gubernator
- Kuber not his
- Cabernets
- Coover Nettie’s
- Kooper Nettie’s
- Cubano
- Kuber daddies
- Kluber Nobby’s
- co-branded
- crew Bernet
Then there’s minikube:
- many Koob
- Mini-Cooper Nettie’s
- mini to
For MySQL it also has some gems:
- Majeski well
- Majeski will
- Maya as well
- my ass Q
- my ask you well
- Eskew all
- might ask you
- my ask Huelva
- my ask Cuil
- Majeski and
- my Eskew L
- Majewski Well
It gives a bit of insight on what’s inside this training set. I’ve even seen the word “nigger” come up in the closed captions where it clearly wasn’t pronounced. Makes you wonder what went in this model…