#273 A Big New Market for Dubbing and Accessibility Solutions with 3Play Media co-CEOs

22:03Christopher Antunes:Tell me more about the script writing. I'm really ignorant about it. So you're saying AI is maybe helpful, but it's not fully kind of used yet. Or like, what's a good script writing look like? And then if you have that original script, is the translation also slightly different? Or is it just kind of just a normal translation done by a pro? Audio description, and we'll talk about translated audio description for global language. Audio description has three components. You have a video that you receive, and then there's three steps. You write a script to describe what's happening, only the relevant things happening in the video for someone who can't see.

22:41Christopher Antunes:So imagine you're watching that movie and you need the scenes described. So script writing, step one. Step two, voice generation. So take that script and translate it to voice. And then three, mixing. So ultimately mix that back with the original video. All three of those steps can be done by people. They can be done by money rounds of people, or they could be done entirely by AI or a mix. I would say most commonly today, we will have an expert write the script, maybe assisted by AI. So we might use an AI draft first in some cases, but ultimately someone is responsible for editing or creating from scratch or approving that script.

23:23Christopher Antunes:And then we'll use voice technology for the voice, and then we'll automate the mix. So two out of three steps, automation or AI. So where's the state of the art in AI script writing? I would say, like in all tasks, the AI is better than people at some things and much, much worse at other things. And one of the things that's worse at is judgment. And we certainly want someone with really good judgment to be the final arbiter of what stays in that description. I mean, we know about hallucinations and things where the AI can sort of go off the rails a bit there. But I think one area where it is very good is in factual identification of things in videos.

24:12Christopher Antunes:The font might be really small and hard to read, or there might be a logo. Imagine someone doesn't know without a lot of research. An AI model can usually detect that thing and describe it quite well. and the judgment part and then we can move on but like the the judgment part would be it would describe things that a human just would know that okay don't describe the tree behind the building

24:33Josh Miller:because nobody cares exactly behind the building it's the pertinence it's the it's the the context and kind of how important is that item or that object or that movement in the scene for someone to understand what's going on just because a car is driving by it's not the fact that it's a car it's the fact that it is the red car that was in the scene before right and just understanding that that's the important part of it is really important.

24:56Christopher Antunes:All right. Let's go to the dubbing you mentioned. I think I checked in Q2 2024, so roughly 18 months ago, you launched your human in the loop dubbing solution. So how has it been past one and a half years? Traction, clients, experiences?

25:13Josh Miller:We've got all of the above. I think we've learned more than we had in the last year and a half. I think we've learned more than we had in the last five years about, you know, this industry, certainly. And just even in product development, I mean, we've, we've, we built a product that I think was really good for some of our customers, but not all of our customers. And I think we, initially, and in the last few months, we've completely changed that to, you know, to really go after what has become a really emerging media demand, which I think is fascinating. So to make that more clear, we have a ton of customers in the e-learning space and the corporate space who are looking at this as an opportunity to distribute content when they had never been able to really localize their content in this type of engaging way.

26:06Josh Miller:What their standard is for good dubbing is completely different from what a media company would consider good dubbing. And a lot of that comes down to their anchoring. What are they used to? Nothing versus something really good. And so that was a very real, both a good thing and a bad thing for how we were building the product.

26:27Christopher Antunes:If you zoom out with this whole dubbing market and kind of what's happening here, on the high end, you have obviously traditional dubbing, theatrical level dubbing, where obviously, Florian, you know, you could be paying$400, $500,$600 per minute to dub something into German or Korean or a language that's difficult. and on the other extreme with full AI and AI dubbing is as a category means a lot of things but AI only with no human intervention at any step you might be paying cents. Our bet is there's a big market in the middle and there's sort of two categories in that market I'd say there and Josh alluded to them.

27:10Christopher Antunes:There's first time dubbers. There's companies, There's whole use cases. There's whole markets that never could afford the economics. We started with talking about unit economics and building a profitable business could never afford 300, 400,$100 a minute, just out of scope. But their brand matters to them and they're not interested in cents. So we're discovering, I'd say together, and I think the market is largely in a discovery phase now on both directions, builders and buyers. We're all discovering together. We're discovering what the right price is. We're discovering what the right quality target is.

27:46Christopher Antunes:And we're discovering how we build that scalably for them. And scalable and repeatable quality is a big topic. You can do it once for a demo. Can you do it a thousand times? Right. And so I think that market is still in the discovery phase. We're figuring out expectations. We're figuring out even the price points because there's a wide range. And what's exciting about 3Play, or for me at 3Play, is the dynamism in this market. it. There are so many use cases, so much configurability, and it's so much more complex. And I love this problem and I can geek out over it for hours. The product and recovery engineer in me will love to do that because it's so complex.

28:25Christopher Antunes:It's not an AI or person choice. There's literally 10 different steps strung together with multiple AI and tooling and expert stops along the way. And I think ultimately that's a big advantage for 3Play because we love complex problems that require orchestration like this. And dubbing is, in order of magnitude, more challenging, I'd say than captioning more description for that reason. And what you said is really true, that a prototype, a demo, a pilot, it's just one thing, but then you want to do it at scale, it all breaks down, right? Even some of the platforms, I log in, I spin up my little video, and then it takes, even that little video takes 30 minutes to render right now.

29:03Christopher Antunes:That's one five minute clip. I guess if you want to do this at scale and 30 language becomes a problem that you seem to like. I struggle with this. I'd love to like, you know, grab a beer someday and dive into this. I struggle with this as across any business. I think demos in a lot of cases, whenever, if there's certainly if there's human in the loop, or if there's expert in the loop in like a model like ours, the whole demo process is broken. What do you learn dubbing one video that's three minutes long into one language. You learn that someone can perform all sorts of magic that's unnatural to produce that three minute video, but you don't learn anything about what the experience will look like in production at any sort of real scale.

29:51Christopher Antunes:So how do you figure that out together in some sort of testing or trial or POC process? That's an unsolved problem. And it's one I think we all need to think about together. Yeah.

30:00Josh Miller:I mean, we saw some of that issue just in captioning, But I agree, it's very much magnified in dubbing when there are so many different steps. And there are really experienced providers out there who know how to throw a lot of people out of problem. And again, going back to our origin, we're not here to throw a lot of people out of problem. We'll throw people out of problem, but in really, really thoughtful ways. And the whole premise here is how do we look at those 10 or 15 tasks that are required to create a good dub and understand one for the customer in the use case, which tasks apply first, and then two, which tasks require human intervention.

30:42Josh Miller:And really, really think that think through that carefully. I mean, we're seeing it now in some of the demos with media companies, and some of them are starting to figure out, and I applaud them for this, how to try to weed out, you know, kind of some of the fake it till you make it, which is, we're not going to tell you what kind of content we're going to send you. And we're going to give it to you on this date and we're going to put a deadline on it, right? And so there's only so much fake, throw a lot of people at it that you can do when you can't really prepare for it.

31:13Christopher Antunes:All right, can we talk about the European Accessibility Act, the EAA, so I don't know, the TLDR, where are we at the various stages of its rollout? I'm based in Europe, but I'm in Switzerland, so I'm slightly detached from the European Union. So tell me a bit more, how big of a deal is it? How does it impact the North American market, if at all? And how much of an opportunity is it for companies like yours?

31:40Josh Miller:It's a great question. I mean, the quick answer is, I think we're in the early days, but we're learning a lot. And so it's certain, we're seeing different approaches from different companies. We started to hear early on from certainly our enterprise clients who have business in Europe, they need to be prepared. And I think for them, it's a little more straightforward, to be honest, it's it's a little easier to say, Okay, we do business here, we have to be accessible, our websites have to be accessible, the videos that support our product launches have to be accessible. Pretty straightforward. That's pretty easy.

32:11Josh Miller:Now, what we don't know yet is the enforcement around that, that's still very early, you know, it just went into effect, we've seen a couple of sightings, but there's not a whole lot of activity yet, because I think there's, you know, there's a reasonable approach here that like, let's give people a chance to get it right and not go too crazy. But I think we'll see soon, more examples being made of companies that ignore it. The media industry is a little bit more diverse, I'd say in their approach. We have a number of companies who we're working with today who've said, we're not screwing this up.

32:49Josh Miller:We're not going to make a mistake. Let's go. We've got to get ready. And we've got to start implementing. We had a couple companies that we were working with literally on day zero. They were ready for the launch of the EAA in June this year. We have other companies that we work with who say, look, we need to know what's possible. We need to be ready, but we're going to wait and see. And a lot of that is that there's some different interpretation of the law itself, that this is where I'm going to say we'd like to play Switzerland and just be not take a side necessarily, but try to be an educator and share what we understand.

33:28Josh Miller:And there are some legitimate ways to interpret the law the way it's written to the point that the lawmakers have actually gone back and said, we messed up, we weren't clear enough. And so what that means is there's some discrepancy around, is it the supporting tools around the media that must be accessible or is it the media itself that needs to be accessible? It is vague enough to interpret either one. Now, a number of media companies have said it's not really up to the EAA to make that decision. It's actually it should fall under the European Media Freedom Act, the EMFA, to make that call. And there are conversations right now about that legislation being updated to kind of take over where the EAA left off.

34:22Josh Miller:But I'd say we're early days. But I think there is a common thread that pretty much every media company is looking at this and trying to figure out how can we be ready.

34:32Christopher Antunes:And Florian, I mean, obviously, like the regulatory environment is more complex because every member nation has their own policies and enforcement practice. And there's a lot to figure out there still. But just like to make it concrete in the localization space, what we're talking about around video that I think still needs to be, you know, ultimately determined. And I think there'll be more guidance coming from the body itself. And I think there'll be enforcement that helps to set precedent. But really, it's if I'm a streaming platform and I've been distributing content in pick any European country and I've been dubbing into 40 languages.

35:08Christopher Antunes:Well, all 40 of those dubs now, in theory, create a requirement to be closed captioned and audio described. That is a significant problem at scale for a lot of companies that have significant distribution in Europe, even if they're North American based.

35:24Josh Miller:That's right. And so what's important there is that this is an unplanned budgeting exercise. And so we go back to what we were talking about before, this idea of synthetic voice being used for multilingual audio description. This is where it's not a square peg round hole anymore. It's square peg square hole because it is so important that they can continue to distribute this content because it's a moneymaker for them and they need to figure out how to do it. And it can only continue to be positive ROI if they have a reasonable approach budget-wise to make it all accessible. To your initial question, Florian, the thing for us that's so interesting is we're still based in North America predominantly.

36:07Josh Miller:We are really working with U.S.-based and Canadian companies who are extending their reach into Europe. we have barely scratched the surface of working directly with European-based companies, which would be the next phase of our growth in terms of the EAA and kind of where that goes. So, you know, we're still early ourselves.

36:27Christopher Antunes:Come to SlaterCon in London. It's London. It's not the EU anymore, but there's a lot of Europeans going to be there. So you guys are a bit of a competitor to ours because you also publish research. You published the state of ASR in 2025 report. report and I looked it up and it says you looked at more than a thousand video files, over 200 hours of content and 1.7 million words. So tell us the key findings of this. So this is something we do, like, you know, and we're definitely not a competitor on the research side. So this is something we did internally. This is a theme with 3Boy. This is something we did internally for a long time.

37:08Christopher Antunes:And we just thought it might be of interest to the market. And we're really biased, but in a really good way. And I'll explain what that means. We consume ASR all day. Like I mentioned, 30 million videos processed, tens of thousands of videos every day. And on the captioning side, our first step is ASR. And then we put that ASR in front of people, and they clean it up. And we are more invested than pretty much anyone in the world. and getting the absolute best ASR so we can speed up that task for people. So we are religious about testing every model. And this is true for all of our human in the loop or expert in the loop services, because the real cost isn't in the AI.

37:54Christopher Antunes:The real cost to us and ultimately to our customers is in the time it takes to get the AI to an acceptable level of quality. So we measure that everywhere and we're religious about driving it down. So we're testing all the engines anyways, so we know which engine to use and in which use case. And so really all we do is take the results of that test and share them with the world because we think it would be of interest. I would say this year, the main finding is that there was a material leap forward, not like an earth shattering one, but there was a material leap forward about a year ago with the generative models entering the mix with the whisper model.

38:32Christopher Antunes:So the LLM or generative AI models entering the mix assembly and whisper most notably. This year, the results kind of plateaued. So we had a big leap forward, but largely the results were the same as they were a year ago. And for English recorded audio, speech recognition is pretty good now. Pretty good. pretty good it's pretty good and i think we all experience that and it it still is highly dependent on scenario so the example i always give which you know i resonates with sports fans at least is if you have you know jim nance at the masters this golf tournament professionally recorded in a studio single speaker that speech recognition is going to be pretty close to perfect um if you're transcribing a focus group for let's say a major cpg company and they've hidden the microphone in a plant somewhere and there's 10 people speaking over each other, the speech recognition is useless.

39:33Christopher Antunes:It's still situation dependent. But for the clean audio single speaker cases, the speech recognition has gotten really good. It just didn't get much better last year.

39:41Josh Miller:Yeah, I think what we're also seeing is, you know, we have the large language model phenomenon, which is great. There's been a lot of talk, and I think appropriately so, about like small language models, right? And so kind of getting more focused on the content itself. And we do this automatically already, the idea of can you build customer-specific kind of vocabulary sets and models and domain-specific models, things like that, that can actually target the type of content you're getting. And I think that's where we're seeing improvements. It's not the broad ASR off the shelf that you can get now.

40:17Josh Miller:So that's where we go back to and believe deeply that ASR is a great tool. AI is a great tool, depending on how it's implemented. And it takes someone pretty sophisticated, I think, to actually get a ton of value out of it, as opposed to some value out of it.

40:33Christopher Antunes:For the AI that you're using generally across the business, for ASR, for dubbing, for all of that, I think you're agnostic, right? Engine, model agnostic, the infrastructure is built, so it doesn't really matter which model leads you can plug in. Yeah. And I think that's, I mean, we thought hard about this. We have a data science team. We have a head of data science who has deep experience in NLP, built speech recognition models in a number of stops before joining. And we thought hard about this, but we really think of it as a feature and not a bug. And I think that's proven true, especially in the last 18 months with the like just massive and rapid improvement and dissemination or propagation of new models.

41:13Christopher Antunes:So the first reason we do it is we, like I said before, all of the cost really here for us is in the people, time and the time on task. And so we need to have the best engine for every task all the time to minimize that time on task to produce the quality we need. And that so we invest in that. Second, because our whole business is starting with AI and sort of cleaning it up, we have tons of data naturally. So we don't have to invest in very expensive evaluation harnesses. We can literally take a new model, plug it in, an A-B test, model A against model B, and we'll get 10 ,000 data points tomorrow.

42:00Christopher Antunes:Because we literally process that much content. So we have like sort of built in rapid learning environment. And we have an ecosystem, a marketplace of thousands and thousands of experts who can essentially do that evaluation for us in a real world context. So it makes a lot of sense in our environment. So let's close on, I don't know, growth strategy. So initially, you said you took on some angel and then, you know, a lot of profitable growth. And then in 2019, you took on some outside investment, I think PE firm, growth PE. What about M &A? like anything in the past and any plans for the future?

42:34Christopher Antunes:What would make sense for you to acquire?

42:36Josh Miller:It's interesting. Literally on day two of working with a private equity firm for us, which was six years ago, they're starting to show you opportunities. So it is built into the process of how do we grow fast, right? So organic or inorganic, their game for anything. So that was a big change for us is to think that way. And so over the years, I mean, we've had plenty of conversations. You know, for us, it needs to be a really good fit. It has to make sense for our initiatives. So there's no perfect answer to say absolutely. Like what we can say is we don't have a mandate to go acquire companies to grow.

43:14Josh Miller:We have a lot of great organic growth in front of us right now with everything we're doing, which we're really excited about. More so than we've seen in a long time with everything that's happening, which is really great. but we're always open-minded. We always look at what we're doing and try to look at the landscape and even lots of companies put themselves up for sale. We are always looking to be thoughtful and if it can add a ton of value, we're going to think about it. If it can't, then that's fine too but it's something that we're always evaluating but it must be one thing or another.

43:49Christopher Antunes:And I would say our sort of growth strategy is not an organic growth strategy where we want to buy a lot of things and that's the way we'll sort of generate revenue and generate profit. But if we see interesting opportunities that align with our roadmap, we're going to look at all them. And as we look out into next year, we're making a lot of investments on the data science side. We're making a lot of investments on localization expertise to surround real localization subject matter experts around this thing we're building so that we can make it as exceptional as we possibly can. That's an area of interest.

44:23Christopher Antunes:there's specific use cases you know we're pretty wide in terms of the use cases we serve for all these products but there's some like that um like you know the the youtube content creator use case is really interesting on the dubbing side now interesting so there's areas like that where there's some companies that are small doing really interesting things that might be interesting targets for us as well i've been waiting for this to like become like almost a mainstream i don't know vertical or horizontal or whatever you want to call it the youtube creator like you know some of them are very big they will eventually want to go international the youtube ai dubs aren't great and uh but the logic is just so obvious like yes we could talk for another two hours just about that this is a this is definitely a focus area for us i mean just just thinking like you can go to youtube and you click and you have 10 audio tracks uh like like on netflix like how obvious can it

45:15Josh Miller:be that and from a content publisher perspective i mean when you think about dubbing and localization at the core of it to monetize the localized content, you have to have a way to distribute that content in another country. What YouTube solves for everyone immediately is distribution. It is a phenomenal concept that people maybe don't even appreciate enough, but literally you have to do nothing to start monetizing other than localizing your content. Whereas we talk with some of our existing customers who say, yeah, we're thinking about expanding into Europe we have to set up a sales team there and sell the content there and find partners there i mean that's gonna be two years before they're ready to actually dub their content whereas with the youtube creator or even a media company publishing on youtube it's it's ready to go it's it's you need

46:06Christopher Antunes:2 000 subscribers i think the thing to solve there and i think you hit the nail on the head like um another another nice feature here is there's a single decision maker it is the content creator are almost always making a call, not a committee of buyers. And they are moving fast, certainly relative to traditional media companies, and they'll take on risk, right? And they'll iterate and they'll learn quickly. And a lot of them operate like almost like at a bigger scale, maybe not the 2000 subscriber ones, but the ones with 20 million operate like little tech, like technology companies. They're rapidly learning.

46:39Christopher Antunes:They're making really bold and quick decisions. I mean, like you said, like if you can prove that this makes sense at a certain quality level and opens up a market so they can monetize it in one language, why not go into 50? There's an economic model to figure out who pays. Do the advertisers pay? Do the content producers pay? Does a vendor like 3Play pay? And there's a licensing arrangement. There's a lot to work out. But I agree with you, Florian. It's definitely a no-brainer and it's coming. So you mentioned a couple of things you're planning for 2026. Anything to add? Any technical cool things you're planning you can you can share yeah so the thing i'm most excited about you know into 26 and as we go out further and like this is more like commentary on the ai economy kind of writ large here that's still that's still emerging but what we're seeing uh with with a lot of our customers i think it's really resonating is um and we think we're going to figure out how to what to name this thing but i'll describe what it is which is our system today like we said it's AI first, and then it's tools and people, and then we sort of deliver those results back.

47:47Christopher Antunes:But what we don't often talk about between AI and tools is analytics and scoring. And so what I mean by that, let's take it in the captioning setting because it's the simplest. Dubbing's a little more complex. But if you have an AI model, and I just mentioned AI can be really good for English captioning if you have a single speaker, we score it. And we score it for internal use. we need to know how much to pay that captioner. If they only have to do a very little to get that to the quality level we need, you don't have to pay as much. If they have to clean up lots and lots and it's going to take them hours, we need to pay them more.

48:24Christopher Antunes:So we've built all these predictive models that tell you how good the AI is and also tell you where the AI fails and where it doesn't. And we use those internally. Well, the realization we've had over the last year or two, as our customers have had macroeconomic climates, made them more budget sensitive, things like that. They would love to see those analytics themselves so they could use them. So in a captioning setting, what this looks like is we run an AI model for you and then we give you a score. And then we give the operational decision and keys over to you. So you decide, is the AI good enough?

48:58Christopher Antunes:Is it not good enough? Do you want us to upgrade? So this concept of AI plus scores that we give you In analytics, we give you to have you make a decision, help you make a good decision about how to allocate your budget and sometimes upgrade with our solution. And even when you upgrade, giving optionality there. So we started this conversation saying there's a big difference between back catalog e-learning and Netflix original theatrical. We call that the dial. So even when you upgrade, you should be able to set the dial exactly where you want to optimize your budget. So shifting this model to AI and then scoring and letting you decide, do you need experts at all?

49:41Christopher Antunes:And if yes, where do you want to set the dial? And doing that across all the services. I would say with captioning today in English, there is a lot of experimentation with just AI. I'll just say. And there is lots of upgrades still. On the dubbing side, it's all expert in the loop. It's all human in the loop for any serious buyers. But we still want to give you the analytics. We still want to give you the scoring so you can see and you can learn. And as it evolves and gets better, we want to be transparent about that with the customers so they can see the evolution of the technology in real time.

50:18Josh Miller:I think that, which I'm very excited about, it applies to even other markets where education, for example, with ADA Title II in the US, they now have to caption and describe way more thousands of hours more content than they're used to this is the only way to do it is is to have more of an ai pass with with intelligence to understand what really matters the other thing i'd say is we are building something that i will i'll speak for myself i'm very excited about in terms of what we're doing with localization dubbing is a great example we've we've not done a great job of telling that story yet and i think the world is going to learn what we've built um this is a great start to that.

50:59Josh Miller:It was a good start. I was going to say it. But we have not really shown everyone what we are capable of in localization, and people don't know who we are as a localization provider still. And I think that's going to change over the next year, and I think people might be surprised what we've been able to build.

51:15Christopher Antunes:All right. Thank you so much for taking the time today. This was a good first step in telling the story. That's right. Thank you, Florian. Thank you for catching up in person. Thank you.

51:27you