#256 YouTube Dub Fail, Propio Buys CyraCom, LSIs Cheer Scale AI Deal

0:00Florian:They're defaulting millions and millions of creators into what is really a substandard AI dubbing experience. You can't edit it. The voice is robotic. Doesn't take turns. No lip syncing.

0:16Florian:Hey everyone, welcome to Slater Pot, an extended version today, extended episode. Hi Esther. Hi Florian. Extended because we have a lot of things to cover. We've been busy with other things and this is the first kind of news episode we're recording in quite a bit. So we are back after, what, five, six weeks and we're freshly baked after a one-week European heat wave. How bad was it in London? Yeah, it was pretty scorching, like 34 degrees Celsius. Celsius. 95 Fahrenheit, I think, around that, 90 to 100 Fahrenheit. So, yeah, crazy. Same here. 35 you know degree sales is pretty bad especially if you don't have aircon and very the northern parts of Europe don't have aircon and I read like the central line gets even 5 degrees hotter.

1:09Well fun fact the Victoria line I think is the hottest line on the tube and that is my line. It's an average of 28 degrees year round so imagine what it is today

1:20Florian:Our brains are baked so pardon if we are stumbling on all the complex topics we're trying to bring to you this week because a lot of stuff has happened, as always, in the past few weeks. But first, you do want to register for SlaterCon Silicon Valley happening on September 3rd and 4th. We have the reception on the evening on Wednesday, September 3rd, and then the full conference on September 4th. We've built a great agenda, Esther. Really, really cool. Starting off with Brian, CEO of Smartling. Then we have a panel on scaling language AI with foundational AI platforms with AWS. So Amazon, we got Google and we got Microsoft on that panel.

2:05Florian:So basically what we're trying to talk about is how do you build all this cool, crazy, demanding new AI on those foundational platforms. And so this is going to be very, very interesting. We have Kohir, Kelly, multilinguality lead from Kohir talking about building multilingual LLMs, really looking forward to that. We have Scott Sukhoff from UC Health. He's running language and cultural services there. That's going to be fantastic. We have a panel with Phrase and then we have the CEO of Sync, a lip sync company, Sync Labs, Prady, and then we have an investor panel that you're going to be running with a couple of investors that are going to be on the website very soon.

2:48Florian:So yeah, and then Ian from XTM is going to talk a little bit to the audience at the network and reception. So you really don't want to miss this. A full day, Menlo Park, be there. All right. Now we did launch our Language Solutions Integrator and Language Tech Platform terms and terminology with the market report. We had a special episode about three to four weeks ago. Now, reception has been very positive. Did you come across some of the reception as well, Esther? I mean, from my side, I think it was great. We've seen a lot of examples of kind of adoption. It's starting to permeate. It's not obviously an instant shift, but, you know, in terms of rolling it out across the industry.

3:35But yeah, definitely sort of once you get into it, talk about it, it's really favorably perceived.

3:41Florian:Yeah, we have, I think RWS used it in their half-year report. So that's a pretty big endorsement. We also received recently an inbound, an M &A kind of, I think it was at the teaser level, right? So, and formerly an LSP is up for sale and they used LSI in the subject, in the email subject line. So, you know, this is great. And also we're going to talk about an M &A transaction in a bit where they use it in actually their press release, the term language solutions integrator. So yes, we're moving. We're moving. This is great. And I think it's very good for the industry to leave the service provider days behind.

4:24Florian:So very briefly, we had a, there was a lot of activity in speech translation. And let me just talk about two kind of different, I think it was three. It was Apple, OpenAI, and then also Google rolled it out for Google Meet. So we don't want to go too much into detail, but it's amazing how much attention speech translation has gotten over the past four or five weeks. So Apple, they had some launch, I think it was their worldwide developer conference or something. Okay. And then you had this person called Leslie Ikamoto, apparently she's director of input experience at Apple. She was standing in this kind of very Apple-y room and was giving this presentation and saying that translation is a feature that Apple is now building across all the communication apps, so iMessage, FaceTime, et cetera.

5:26Florian:And the kicker is they said that it's all run on device, which seems tricky, tough. I mean, run on device means it's also probably from a latency point of view, it's much easier to do. But still very tough. And then, but the demo was very basic. It's like, hey, you talk to your grandmother and it's Spanish and it's kind of like taking turns and it's all clear and so on and so forth. So yeah, let's see how successful that's going to be. Then OpenAI rolled out improvements to their advanced voice mode, which is a feature for paid subscribers, promising more natural human-like interactions and also real-time AI speech translation.

6:08Florian:You remember that Italian demo that they gave back in 2025 with the CTO that's no longer the CTO because half the company has left, I guess. Well, they're still hiring a ton of people, of course. Yeah, it's certainly hard to keep up on the hiring side, personnel side, as well as the tech side. We're going to talk about Meta later, but Meta has just poached a bunch of OpenAI researchers, apparently giving them like$100 million as a signing bonus, including a couple of researchers in Zurich. Anyway, back to advanced voice mode. So they're promising new and better real-time speech translation capability.

6:41Florian:We all know how hard it is. So I think we tested it briefly, found it only works on the mobile interface, that it kind of switches between language learner mode and conversation mode, even though you're prompted for translation. So the AI is a little confused what it's supposed to do. And then sometimes it apparently kind of stayed silent, didn't translate. Now, you know, let's see. Those are all things that are in the early, kind of in the early innings. And yeah, they're gonna evolve. So have you come across any weird YouTube AI dubs the past couple of weeks, Hester? Not that I've noticed. No.

7:20Florian:Okay. I have. Yes. Okay. Tell me more. I want to talk about YouTube dubs. So this is, it's a very, I think it's a very big deal. So they announced that they were rolling out this AI dubbing feature, which we must have spoken about many times in the past on this. But now it's getting like real. So they rolled it out to 80 million creators, which must be like almost all of the creators. And they expanded language coverage to 20 languages. so that AI dubbing feature originally released in 2024,

7:58Florian:you needed to qualify for it. But yeah, now it's kind of widely available. Yeah. And we've spoken about this for so long and I always wonder like why are they holding back? Why are they not launching this? They could have launched this two years ago maybe but now I know because it's really bad. I mean, it's really bad. Okay. So currently it's enabled for almost every channel, right? Yeah. And then for those on YouTube, I'll do a quick screen share. So the voice is extremely robotic. The translation is actually pretty bad as well. So just in one test that it actually mistranslated literally the term AI into German as like, just, yeah, mistranslated it.

8:51Florian:Yeah. Then I came across another video just before the podcast where it actually translated a pretty bad swear word, like expletive in, it was pretty bad in English, but it was way worse in German. Oh, over translation. Well, but it's a pretty bad word. Yeah. So I'm not going to say this is a marginally family-friendly podcast. But the point is, like, it's one thing if your creators upload things that are, you know, explicit or swear words or things like that. But if then Google goes and uses AI to make that word even worse in the target language, I don't know. It feels a bit dicey. And there's currently no way to edit the AI dubs.

9:38Florian:I think only some very big creators have this. And that's what everybody is complaining about also on the forums and in the comments, right? And there's also no way to upload your own AI dubs. So you couldn't use a third-party AI dubbing app and then upload your own dubs. Only the top creators can do this. And so anecdotal, I think to me, evidence is like most careful creators have opted out so far. I haven't seen much. So let me, for those of you who are on YouTube and the others just, you know, do go on to the YouTube. I'll quickly share the screen so I can show you what I mean here. So this is what we get, we being our Slater YouTube channel, right?

10:27Florian:We have all these options now and it basically says publish, but it's in draft mode. Right? So you can go in, so I'm looking at a list of a bunch of, um, a language. I think this must be about 10 or something, maybe the standard one. So I'm going into, for example, German, right? And then I have the option here of choosing the original or going with the AI dubbed German, right? It already, uh, the show notes, it already translated more or less. Okay. Ish. Right. And then I can go in here and I can click.

11:09Florian:Okay, so if you've heard this, it's pretty robotic. It's the same. And this is a podcast that was our most recent podcast, right?

11:26Florian:So it uses the same robotic voice. It doesn't take turns, right? It's the same voice for both speakers. Yeah, okay. And so it's really painful to listen to. Plus there's no way to edit. I can't edit it. So who's going to go and say, well, sure, let me AI publish this and the AI adopt version, even though I have maybe a couple of things I want to fix. So I can't fix it yet. Yeah. Not great. And I mean, if look at this one, this is actually YouTube's, I'm sharing now the actual launch video by YouTube itself. And you can listen to this now. So if it's Tuesday, it's Tuesday. So let's get into it.

12:10Florian:It's Tuesday. It's Tuesday. So here you can now change the audio track, like on audio track, right? And you go and select German and off you go. So, let's start. We're going to make 80 million creators to make the world's automatic synchronization. Creator? I mean, we don't have to go and listen much more to just hear that this is not great. It's not great quality. It's also kind of a robotic voice. Yeah. So, not sure why this launch. like they've had three years and now they're coming out and basically they're defaulting millions and millions of creators into what is really a substandard AI dubbing experience.

12:58Florian:You can't edit it. The voice is robotic. It doesn't take turns. No lip syncing. It mistranslates. It translates swear words. I don't know. Okay. Anybody at YouTube listening to this? They got some work to do. You got some work to do because this is big. This is not just a feature. This is like you're putting this into, there's millions and millions of hours now, I guess, being like AI dubbed and translated of people that probably don't care about language as much as we do. And, you know, they have this now on their channel. So like, why? All right. Like we keep with the topic AI dubbing because in RWS's results recently, half-year results, they announced that they acquired the IP of Papercup and they're going to integrate it into Trata.

13:49Florian:So Esther, tell us more about this. Yeah, exactly. I think you hit the nail on the head though. It's not the entire company Papercup. It is just the IP of Papercup that RWS has acquired. So Papercup being an AI dubbing company, Papercup based in the UK, RWS also based in the UK. we had the CEO of Papercup on our podcast many, many years ago, what it feels like many years ago now. I think we had some coverage a couple of weeks ago, it must have been now, where we pointed out that the team, or much of the team from Papercup, had joined Scale AI. So as we said, this is sort of a partial acquisition by RWS, not of the entire company, but just of the IP.

14:34from RWS's perspective, I mean, I think this is clearly forming part of a broader plan to focus firstly on AI capabilities and then also kind of centering on multilingual video, voice, and the kind of content localization piece generally across its tech stack, like you said, with Trados. So yeah, very interesting one.

14:57Florian:I mean, if you integrate this into Trados, You know, technology is one thing, but distribution is another. So you can't get much more distribution in the language industry than with Trados. So to me, it seems like a, on paper, very, very smart move. No pun intended. Yeah. And integrated. Yeah. Yeah. Also in those results, RWS announced, I mean, we had the new, the new kind of corporate identity, new logo already, but they also announced the new simplified go to market positioning, which I also like. So they're going to reorganize the group into three business units called Generate, so that's content tech and the train AI business, so the data labeling business, then transform localization, including all the language, tech and services, and then Protect, which is the IP services business.

15:51Florian:So Generate, Transform and Protect. And they're going to also report along these three business lines. So they're going to change the reporting. So that's good. Yeah. So I think this is a pretty smart mood. Makes it much easier. We don't have to go through the results too much. It's kind of very stable. Revenues are down marginally to like 344 million pounds, half year that is. And then profit is down more, but there's some accounting stuff going on that I can't unpack here. Divisional performance. I think that the tech business is doing very well with Propylon, LanguageWeaver, and the SaaS. Software as a service transformation for Trotus is moving quite well.

16:35Florian:You know, IP services apparently declined, but you know, not a real expert on that business. And then regulated industries also declined to the operational challenges, they said. The IP though, they divested something, PAP-based, so that could explain some of the declines there. Yeah, good catch. Yeah, and in terms of the outlook, modest organic growth, low single-digit growth, and pre-tax profit remains at 60 to 70 million pounds. So that's a very profitable business. And then one of the businesses that's doing really well, they pointed out, is Train AI, which must be benefiting from what's kind of a tectonic shifts in the data labeling business currently.

17:19Florian:Yeah. Because, and let me try to unpack this. Let's try to unpack this together. This is a complicated one. Yeah. So, ScaleAI, the leader in data labeling and a key competitor to a bunch of the LSIs, the language solutions integrators that we know, like the TrainAI from RWS and DataForce from TransPerfect and what have you. So, Scale.ai received a major investment by Meta and so Meta invested$14 billion into Scale and it bought 49 % of the company. And the founder of Scale.ai joined Meta now. I think his title is like Chief AI Officer or something. Very nice. So I'm not sure how that's going to resonate with Jan LeCun who's the chief AI scientist, but anyway.

18:22Florian:And as we said before, they're poaching a bunch of people from OpenAI, but let's go back to the topic at hand here.

18:32Florian:So scale kind of sold in a sense to Meta or at least a big part of it or half of it. Almost tough, yeah. And the Papercup team also went to scale. So, the Papercup team is kind of now part of like a kind of a new meta operation there. So, now, I mean, the current CEO or the interim CEO of Scale said, hey, hey, you know, everybody relax. Like, you know, fewer than a dozen people left and went to meta, we're servicing our existing clients and yada, yada, yada. But I mean, you have a lot of other voices out there that are saying there's basically now a free-for-all because all of the scale AI data labeling clients or many of them are direct competitors of Meta and don't feel really comfortable being served by a company that's half owned by one of their biggest competitors.

19:29Florian:So there was a great episode of This Week in Startups where they had, among other people, Manu Sharma, the CEO of LabelBox on the podcast. And that used to be kind of a tech only solution for data labeling. And then he talked about how they also went into services and then really growth took off because probably just the tech is maxing out at some point. So you want to go into the services component and they're seeing a lot of customers from scale coming over to them now. So I'm trying to onboard them. So this is a huge opportunity again for RWS with TrainAI for WeLocalize. I think they have VLO data that's doing really well.

20:15Florian:Then Somalingua, we had Veronique on the podcast in that space, right? Appen, of course, or a big opportunity also for TransPerfect, their data force unit and then Alcolot. Alcolot is in that business. So I think they did a couple of strategic hires. We picked that one up. So maybe can you tell us a bit about that? So the press release was essentially talking about expansion of its AI data services solutions. So very much focused in on the topic we've just been discussing. So kind of scaling out data collection, annotation, all those multilingual generation rewriting etc and to do so they have appointed at least announced a couple of appointments there in so one person in North America that's appointed as business development director for AI data services that's Sang Do Lee who has joined to lead AI and data growth strategy for North America person I mean as we've just been saying the person has experience with Appen previously.

21:18Also, I think a few years ago, experience with RWS Moravia, as well as a company called Defined AI, which says it's kind of focused around ethical AI data and a marketplace for sort of data access and custom solutions. So very much coming from within that sort of data for AI space. And they also appointed somebody called Melina Carullo as head of production for the AI data division. She comes from eight years at E2F.

21:51Florian:Quite the poach. Okay. Yeah, across data and AI solutions for E2F previously and account and program management. So yeah, clearly like two people that are very knowledgeable and experienced within that market. And again, it's the same sort of names and competitors that you're hearing from here. Well, I mean, now is the time to double down on this. If there's millions and millions of revenue from scale to, I don't know, scrape up. So scale reported, how much revenue? They reported$870 million in revenue in 2024. Well, that's a lot of business for you to go after if you're in the space. You could probably extrapolate for the valuation as well, couldn't you?

22:42If you're quick enough.

22:44Florian:Remember that story with Appen, where Appen had to issue a profit warning because they lost the Google account? And the Google account maybe went to scale. And so now the last thing that Google wants is having Meta label their data. Correctly. Yeah, correct. Own half the company that is responsible for labeling their data. Yeah, it's, I mean, you can't think about it too much because you'll end up going round in circles. From RWS to Scale to Papercup to Meta to Google to Appen and back again. So, yeah, very confusing. I wanted well in this though. Why 49 %? Is that because of some like anti-competitive?

23:24Florian:Well, the controlling thing, right? You can't have 50%. Equally, they don't want to sell the majority of the company. I think they wanted to really sell the majority of the company. I think this must be some regulatory issue because there had been all these, I think the big tech companies were kind of essentially banned from buying, from doing M &A. That was a big topic in the VC community for years now where, what was her name? Lina Khan. I think she was, I'm blanking on the title, but I think she made it very hard for big tech to buy, acquire a company because of like some anti-competitive concerns.

24:03Florian:So yeah, maybe that's why they... But this is as close as you can get to a controlling stake without actually being a controlling stake, right? Yeah. So maybe that's what they did. Hence the concern of all the big tech competitors about having Meta basically labeling their data. Another competitor, and I do want to mention it because it's fun. I like Grok, the XAI chatbot or whatever it is. They call it Grok. So they ditched Google Translate because it used to be that on Twitter. And then later on X, they had Google Translate as the default translation. If you want to have a tweet or an X post translated, you could click and then it was Google Translate and now they switched it and they do it in-house now.

24:48Florian:So Grok's now translating tweets or X posts. And yeah, users aren't happy. Because one thing is that in the past with Google, it was like you clicked and you saw the original and then you saw the translation and now it's like overriding the original. Okay. And it's quite slow. So it's like AI speed translation. So it's kind of slowly, slowly overriding it. Yeah. Well, welcome, Grok, to the translation game. All right. Big news in M &A and funding. We have three acquisitions to talk through, and then we want to close in a couple of big hires. Yeah. Well, so first of all, we have a big breaking news with a deal that closed on the 1st of July.

25:34That is the acquisition of Siracom by Propio. So US to US deal there. No financial metrics are being disclosed, and that's fairly standard. But yeah, you've got the acquisition of a very large interpreting provider in Syracom by a very large, also predominantly interpreting or at least sort of traditionally interpreting focused language solutions integrator in Propio to kind of come together combined to be probably at least half a billion dollars in revenues. the combined organization there. So this obviously follows Propio's prior acquisitions. In 2024 alone, they acquired ULG, Acorbi, and a company called ASL Services, already adding tens of millions of revenue to their top line there.

26:41So really kind of going through this hyperscaling in 2024 up to, I think they were fifth on the Slater LSPI index at sort of plus 300 million. So they're already a super agency in 2024. And now here buying what was the largest privately owned interpreting provider, I believe, in the world.