I asked my 13yo to watch #CrunchLabs on Netflix with me. I had read posts about its #AIdubbing and I was very curious about his reaction. So we sat comfortably and I went to get some snacks. By the time I got back he had switched it to the original voice with Greek subtitles.

Some background is necessary here: my son is dyslexic and he doesn’t like reading, so he normally prefers dubs to subs. Plus, his English is intermediate, certainly not enough to understand the scientific terms in the show.

So I was really surprised by his choice. I tried to nudge him back to the dub (for my research purposes, of course). He was adamant: “It’s stupid.” I asked why, hoping there was more nuance behind the “stupid”, and got: “It’s fake and I can’t stand it.”

That was another surprise. He spends a lot of time like most kids his age, watching YouTube shorts with synthetic voices. I honestly thought he wouldn’t mind, and I wonder why he did.

Anyway, we watched an episode together, it was fun, and then he went off to play Fortnite with his mates.

I switched to the dubbing.

It is interesting that Netflix is experimenting with this. The AI dubs are the same voices you hear on Mark Rober’s CrunchLabs YouTube channel
– and he’s already been mentioned as one of the first creators testing YouTube’s AI-powered multi-language audio to grow his international audience (for instance in this TechCrunch piece https://lnkd.in/gcwvjw6X).

Netflix’s own researchers have been studying how to evaluate AI dubbing and lip-sync quality, for example in this paper on audio-visual lip-sync estimation for dubbed audio https://lnkd.in/g5e5Yg2B.

I don’t think any evaluation technique was used in this case. Not because of the unnatural pauses and odd intonation (never mind the translation errors), but because there was so much overlapping that in some segments you couldn’t understand what anyone was saying.

Then there are the rumours about Netflix creating their own AI dubbing technology called “DeepSpeak”. As far as I can tell, they are still unfounded. The product name seems to first appear in an article by Eid Mohamed, which doesn’t cite any sources (and he never replied when contacted). That article was then picked up by a student for a case study published by The Localization Institute. The student paper is clearly AI-assisted (OK, it's AI slop). Some outlets then cited *that* without going back to check where the claims actually came from.

In total, trying to dig through this felt like a bit of a waste of time – but also a small case study in how quickly misinformation and unverified narratives can spread, even in our little corner of localization and media tech.

YouTube’s multi-language audio feature for dubbing videos rolls out to all creators | TechCrunch