adventures in autocaptioning

[kaberett]

Today I finally got around to watching through a workshop on rare earth elements I'd signed up for and then slept through in February (and it took me a day to manage a half-day* workshop, but I'm pretty sure I got more out of it than I'd have managed live, so there we go).

Obviously, given the option, I turned on autocaptioning.

Ways that autocarrot rendered "rare earth(s)", in the context of "rare earths group", "rare earth patterns", and the like, over the course of the workshop, in order of appearance:

rabbit elements, rarest elements, relative element patterns, reverse grip, arrest elements,various elements, various, arrows patterns, rarer, red earth, "two quick ways of looking uh using railways", light rail, rare settlement, growth patterns, elephant patterns, relevant pattern, earthling patterns, flame pattern, real thermal pattern generator, real-time pattern, retinal pattern, wealthtrack patterns, weapon patterns, offline patterns, earthern pattern, real filament

* i.e. nominally ~2h of recorded material

Comments

[batrachian]

Oh, technology. *sigh of amused exasperation*

[serpentrose]

Yeah, I've more than once had a laugh at autocaptions.

[alexseanchai]

lololol

[rugessnome]

oh dear. One food youtuber I like to watch (emmymade) has a habit of saying itadakimasu before most taste tests and I believe there's a running joke at her channel about how the expecting-English autocaptioning renders this in various ways, to the point that iirc she sold (and had) a "Eat the ducky moss" shirt at one point.

[serpentrose]

heh, that's awesome. I want that shirt. (Though I'm not allowing myself to buy clothes until further notice.)

[vass]

Elephant patterns.

I think an elephant might have done a better job with the captions.

[judiff]

I want Rabbit Elements to be a thing

[gumbie_cat]

At one point, part of my job involved correcting autocaptions and I used to keep a text file of all my favourites. The best one is still that the place I work frequently got rendered as the "University of Not Again!"

[kaberett]

HA.

[silveradept]

That's a good list of wrong possibilities, so I'm pretty sure we don't have to worry about the machines taking over any time soon.

[hilarita]

Sigh. Accessibility fails again (still).

Autocaptions are pretty terrible for most research domains, though they may not suck so much for computer science...

[kaberett]

The thing is, I know e.g. Dragon has developed specialised vocabularies for specific fields, so really if a thing has "Rare Earth Elements" in the text title one sort of... expects... our Googly Overlord to be able to make reasonable inferences from there???

[hilarita]

I don't think the probability model is that context sensitive.

[jesse_the_k]

This is a very useful collection of data, thank you.

Text-to-speech is the most evolved of these computer-attempts-language tools. The great thing about TTS is it's consistent. If your name is Rocinante Cischester the TTS will mangle, but it's the same mangle each time.

Speech-to-text, whether it's an input method replacing the keyboard or autocraptioning, is working with our variable human voices. The results can be widely different, depending on word placement in a sentence.

The most irritating thing is lack of speaker ID, because that's something we humans can fix by routinely announcing our names as we begin to speak.

"This is Prof Partridge, disagreeing with your earlier analysis"

[glass_icarus]

Ugly laughing, wowwwwww.