


20 Nov 2026 · 44m · 20 chapters

Iyuno CEO David Lee discusses how AI should transform media localization (subtitling, dubbing, dialogue lists, audio description, compliance edits) without removing humans, focusing on Iyuno’s “Chloe” contextual memory engine.
David Lee is CEO of Iyuno, founded in 2002 in Seoul by three founders (Iyuno from surnames Lee/Yoon/No). He studied civil engineering, built early translation/timecoding software, scaled Iyuno in Korea, then moved to Singapore and later expanded to the US/LA during Netflix/OTT growth. Iyuno employs ~2,000 people and works with tens of thousands of freelance translators/voice talents across 34 countries.
AI is shifting from “vendor-push” to client-driven transformation. Chloe captures “story once” context for consistent multi-output localization; humans remain essential for QC and creative judgment. Iyuno uses frontier multimodal models (no in-house model training) plus specialized agents and graph memory. Voice AI requires consent, control, and compensation (“3 Cs”).
Chloe boundaries can span episodes/seasons/franchises (e.g., Avengers vs Harry Potter). SAG-AFTRA consent/residuals approach. AI micro-dramas in China use AI-generated content but still need human dubbing due to budget and controllability limits.
Written by AI. May contain mistakes.
David Lee shares his background and the founding story of IUNO.
“Hi, David, and thanks so much for joining us today.”
Discussion on IUNO's growth, acquisitions, and current status in the industry.
“And up to the point where I decided how to move up and expand.”
David discusses the impact of AI on localization and IUNO's strategy.
“And when the models, the LLM, the large models came out, it didn't really take long for us to realize that model's not probably the issue.”
Introduction and overview of Chloe, IUNO's innovative AI engine.
“It used to be the latter where we were pushing for like, we have this great solution called Chloe, it's going to make all these things possible.”
Exploring how Chloe retains context for better localization outcomes.
“But where we saw the problem was really context because we had that experience of the neural machine translation days.”
Discussion on the evolving role of project managers in localization with Chloe.
“So how does Chloe gather and retain context in terms of, I guess, the context window?”
Learn how Chloe simplifies the localization process for projects.
“Now with Chloe, that just simplifies and it just compacts in terms of turnaround time, so it just simplifies that workload for them as I see it.”
Discover the impact of AI models on cost and efficiency in localization.
“So by the way, we don't develop any models.”
Explore how Chloe is being used in subtitling and client projects.
“And then on top of that, we're putting a lot of inference across a lot of data points.”
Understand the importance of consent in AI voice dubbing practices.
“pretty rapidly, both internally and externally, I think getting the right traction as well.”
Discuss the role of AI and human talent in future media localization.
“We're, we're sort of moderating in between the clients and the voice cons in terms of the AORs.”
Learn about the challenges and limitations of AI voice synthesis technology.
“I think we'll sort of follow the trend, but then we'll react as we go.”
Exploring the difficulties AI faces in accurately synthesizing voice based on emotions.
“If there's no sample for that particular emotion, the AI is struggling to create that.”
Discussion on the emergence and profitability of AI-generated micro dramas in China.
“So I think those days will come pretty soon.”
Understanding why AI-generated content still requires human dubbing despite advancements.
“Is there any opportunity for like, you know, kind of premium dubbing companies if this catches on or maybe doing business in China from things like that?”
Analyzing opportunities for consolidation in the localization industry and future directions.
“narrower localization space, or would you be looking at kind of adjacent and acquisitions in technology or maybe other kind of adjacent production services?”
Insights into the development and implementation of the Chloe platform for localization.
“Is the primary channel for that mostly YouTube or what other kind of distribution channels do you see for that type of content?”
Discussing the necessity of human involvement in media localization and the challenges ahead.
“At the same time, obviously we'll be looking at adjacent markets to grow the business.”
Explore the complexities of sourcing quality audio data for AI model training.
“So just the sheer amount of data out there to train a good model has always been the constraint when we think about audio models or video models.”
Discuss the balance between automation and human creativity in media localization.
“I think that's going to take some time to get there.”

Get the gist, explore an idea or ask about something said.
On SlatorPod, what are the key points of this episode?
Automatic transcript. May contain errors.
0:00We're not really looking at a full automation, whether it's, you know, subtitling or whether it's project management, whether it's dubbing, we still believe, and I still believe humans are very important piece of that loop.
0:16All right. Welcome everyone to another episode of SlaterPod on the podcast today. We welcome David Lee, the CEO of IUNO. So IUNO, for those few of you who don't know, is the world's largest media localization company. So if you've watched any dubbed or subtitled show movie over the past few months or years, chances are very high you've seen Iyuna at work. David gave a super interesting presentation at SlaterCon San Francisco in person in September. And so I'm very happy to have him on the podcast. Hi, David, and thanks so much for joining us today. Hey, Florian. Nice to see you here in this very renowned podcast.
0:54I've watched a lot of your past podcasts, by the way, so happy to be here. and thanks for having me. Thank you. Well, we're at episode, this is 302 or something, so yeah, we keep up the pace. Where in the world are you recording this from today? I'm not in LA. Yeah, today was a back-to-back call, so still at my home, finding time where I could commute to the office, but it's what we live these days. Yeah, that's the right place to be if you're in that business. All right, look, let's start with you telling us a bit more about your professional background, journey, how you got started with IUNO.
1:34I really don't have a background per se, to be honest, because I started the company when I was in my last year in university, 2002 out of Seoul. It was three founders and some of you know the story, but where the IUNO comes from is really the surnames of these last names of these three founders. so in Korean my last name Lee is Lee and then there was another person last name Yoon and then there's another person last name No so combining those three was was I you know but it was we started off as a as a translation company we initially did a lot of subtitling work for DVDs back then we had to use VHS tapes eventually these contents became DVDs but we we use VHS tapes and word processors and I thought that was very inefficient so back then I was well by the way in college I was majoring civil engineering but just decided like this is really inefficient we need to build software to do this so back then we I hired this friend who had developed this software that could do you know file level you upload the video files and do time coding and translation so that was really the start of okay let's try to solve this you know problem with technology sort of had that mentality back then.
2:52And then spent about eight years trying to bring up the business in Korea, there was some trials and errors, a bit of success, a bit of failure as well. And up to the point where I decided how to move up and expand. So moved to Singapore, where back then, a lot of the regional broadcasters were out there. So moved over there and then managed to grow the business to a meaningful size out of Asia. And because back then, you know, the experience of, okay, global expansion is nothing more than just buying a plane ticket and going there and knocking on the doors and seeing for yourself what's going to happen.
3:31That pretty much was the case when I decided like when Netflix was coming up in OTT and all this, you know, transition was happening, we just have to be, you know, in center of gravity, just bought a plane ticket again and flew over to LA and it just expanded like that. So again, it's language, but we always like, I personally always like to use technology to address problems. So that has been the mentality. Yeah. I love how you stopped at, you go to LA when the whole OTT and Netflix and streaming boom happened. So let's take it from there, major acquisitions. Now we're in the AI boom. So kind of give us an update of where Iuna stands today and the main priorities for the business.
4:17Sure thing. Well, obviously we've become the largest and after moving to US and sort of taking off, we introduced some funding capital to help us grow the business. So we were acquisitive and we merged with two different companies, BTI and STI over the past years, pre and during COVID. We're now at about 2 ,000 people globally in terms of full-time employees. And we work with tens of thousands of freelance translators, as well as voice talents, i.e. creative talents, across 34 countries where we operate. and really the initial part of the merging, post-merging of these companies, really our effort was to bring the companies together just to really culturally and technically and operationally become one team.
5:12And I think that's been paying off for the past years. But now that we have AI in our way, obviously any localization company, we've met up with a lot of the other adjacent localization companies during SlaterCon as well, but AI is a very main topic, not just for, by the way, localization companies, but for the content creators as well. And I think we've spent a number of years to set out the right vision for our future when it comes to the use of AI and how that transforms us and how we need to adapt and what sort of is our new target operating model and how we then interact with the clients, how do we service the clients and all that, right?
5:53And I think the opportunity we had before the LLMs came out was we had spun off this neural machine translation company. We had tested out AI back then. So we knew what the limitation was. And when the models, the LLM, the large models came out, it didn't really take long for us to realize that model's not probably the issue. So we built this contextual memory, which I guess we can talk in a little bit more detail. But, you know, obviously our priority right now is to really use Chloe to, first of all, make ourselves very efficient by leveraging that contextual memory. But then also, you know, servicing our customers better because they are looking for innovations both internally and externally.
6:40And they are looking for partners that can effectively deliver more subtitles, more dubbing for them in a more cost efficient way, but with better quality, by the way. but in a much faster turnaround time given how their clients and viewers, us in general, are evolving towards, you know, we have much more content to watch. They're just exposed to much more competition themselves. So they need us to adapt to that new dynamic. So that's what we're really trying to do here. But in all parts of this, which is sort of the second most important priority for us would be our people and our creative talent network and our collaborators.
7:22So in any of our AI approaches, we're not really looking at a full automation, whether it's subtitling or whether it's project management, whether it's dubbing. We still believe, and I still believe, humans are a very important piece of that loop. We call it human in the loop. whether it's QC approvals, judgments, creative inputs or directions, our strategies that are creative talents and the trust especially we've built with these creative networks and talents for the past decades come along with our future alongside with our AI strategies. Are you seeing the push? Is there a big push from the clients, from the big clients to kind of become, I guess, more AI-centric or do you feel sometimes that you would actually have a slightly more or a more AI-centric strategy and the clients are a little more hesitant at this point?
8:23Like what's the kind of balance there? It used to be the latter where we were pushing for like, we have this great solution called Chloe, it's going to make all these things possible. Do you want to bite at it? and that was maybe two years ago, maybe a year ago, but over the past six to 12 months, things have really, really rapidly been shifting. It seems like, by the way, our clients have been really going through a lot of changes, by the way, leadership changes, consolidations, as we know, and so on. And they're really effectively bringing AI into all parts of their business. I think as content creators, obviously they themselves need to leverage this technology technology to meet up with the game, with the competition, like I said.
9:10But yeah, it's really the other way around nowadays. They're pushing for transformation. So you did talk about Chloe. So I think Chloe stands for contextual life cycle orchestration engine. Yeah. So give us Chloe in a nutshell, and then I want to dwell on it a lot more because you gave a super interesting presentation at Slatercon about it, and I really want to unpack this for people that weren't part of that conference. Sure thing. Like I said, I mean, for years, we've realized the models were initially the problem because we weren't getting grammatically proper translations back then during what Google Translate or the statistical translation days.
9:52Now, obviously, we've passed that, right? The models are there. And it was our belief that the models are going to get better and better. I think over the past year, it's just significantly nobody disagrees to that. But where we saw the problem was really context because we had that experience of the neural machine translation days. And think of hiring a brilliant translator but not giving them the video file in the first place and just pure script to watch to translate. And that really doesn't bring out the best translation because they can't see the video. They can't imagine the scenes. They can't hear the nuances and tone and manner of the characters and so on.
10:34So it's like if you don't give them the context, there's just no way a relevant or accurate or even creative outcomes would come out. The other problem is repetition. If you're doing subtitles, if you're doing dubbing, if you're doing audio descriptions, if you're doing compliance editing, Truth is everybody's actually watching the video repetitively, probably interpreting the video in their own different ways. Maybe that's not good. So every version is starting from zero. So that repetition is another problem sort of statement we saw in those old days. So we saw an opportunity where, you know, Chloe would flip that actually in a way that we would capture the story once, i.e.
11:18we would build an agent. We would build an engine that would actually watch the content on behalf of our humans. And probably not perfect, but up to a very, very accurate level. And then once it's built, that single source of truth where all the context lies could be activated anywhere. So then you could export that into subtitles. You could export that into dubbing scripts. You could export that into audio description files and the scripts or like compliance editing EDL files and so on and so forth. So each skills based on the same understanding for, you know, to generate these results. And then adding human as sort of the check layer and shape the results was really the strategy behind Chloe.
12:05So how does Chloe gather and retain context in terms of, I guess, the context window? Is it like episodes, seasons, entire franchises, things that I guess like Breaking Bad is, I don't know, it's like 10 seasons, right? Yeah. It really depends on what the content owner's intent is, right? The boundaries for that context, right? Obviously the Avengers universe probably has to be separate from the Harry Potter universe, I guess. They don't mingle, right? They're just separate. They don't mingle. Nope. Yeah, exactly. Yeah. So, I mean, if we have sort of a universe defined, and that could be multiple seasons of episodes, if not different franchises could all be part of a single universe.
12:53There are several, you know, universes that have different franchises underneath. So it really depends on how you want to build that. But obviously we have episode by episode sort of boundaries as well. So that, you know, obviously we could refer if we're doing a subtitling for the specific episode, you know, the contextual memory, the agents would dig into the other episodes for context that it's not sure if it's this is whether it's right or wrong. Or was there the right KMP key names or places or terminologies used for this character that might have appeared in a previous season and so on and so forth.
13:34So yeah, the boundary goes depending on how you want to define it. What kind of person, what kind of job profile sets that type of boundaries on your side? Like what role? Because that role didn't exist four or five years ago, right? So like, who does that? right yeah if we get a series from a client hey mr client wants us to do a season of 10 episodes of a season there's definitely a project manager within us who would go into our tools and select season and episodes and then upload the files and distribute it to our workforces to either localized i mean subtitle or dove so that project manager still stays so that project manager would define the boundaries of okay this is the new season that's came in that's come in we actually did a previous season so it should be tagged under the same say universe or franchise so that project manager doesn't go away they probably just find it easier to work on chloe because in the past they had to wake wait for or distribute it to like all these different type of people or you know pipelines lines and wait for what, two weeks, four weeks, or dubbing could take up to what, 12 weeks for them to receive the results and track that.
14:56Now with Chloe, that just simplifies and it just compacts in terms of turnaround time, so it just simplifies that workload for them as I see it. Not sure if I answered your question correctly, but. Makes a lot of sense. So where does the, I guess you're probably model agnostic, I guess, in terms of, You know, obviously we're not doing our own NLMs anymore and all that, but like where does the model get the kind of the base layer intelligence from that now you can then add the additional context to for these individual projects? So by the way, we don't develop any models. When we initially kicked off Chloe, there was a clear gap between sort of the offline, we call it open source models that we could fine tune versus the frontier models and obviously the big companies would not.
15:47release these open models they would always focus on the frontier models as sort of um in terms of their best um best technologies um so so we don't use any of our own models we use frontier models and they they're very good and they're improving continuously improving in in in extracting the raw language and and the vision abilities right are i.e if we think about human are our eyes and ears um so so we use those foundation frontier models to extract those you know what we call the you know first layer of the sensory layer of information once once those are extracted i.e the time codes and dialogues and all the objects and events that happen visual and audible events then we've where where our moat is we've designed these specialized agents that infer by connecting the dots and building the story narratives, building the character personas, oftentimes drawing the blueprints of a certain premise so that we could infer where this person has actually entered when opening that doorknob and where he's entered and how that connects.
17:00Because that's what we do in our brain when we watch content anyway. So we try to mimic how we and reflect on ourselves in terms of how we watch content. And once those data points, we call it artifacts, are then inferred or built, then we structure that into a graph memory so that we have the relative relationships that we could track. And so that effectively AI models, when we call it skills, but when we want to output to a skill, i.e. could be subtitles or AD scripts, like I said, what we want those outputs, then it could refer to that graph memory to bring the right information artifacts out so that it could make the right translations or iterations or whatsoever.
17:48That makes sense. Sounds very compute intensive. Is that true? I mean, is this cost an issue in this? Because especially with vision, I mean, I guess it gets expensive quickly. It is. But the fact that we could only do it if you, I mean, in the traditional. So it is very expensive in the sense that we're using very expensive models. First of all, the frontier models, as well as the vision models and the audio model. Multimodal models are expensive. And then on top of that, we're putting a lot of inference across a lot of data points. So yes, it's pretty, it's not as cheap, but at the end of the day, it's not as expensive either.
18:26But when we think about in the traditional world where human had to watch all that, given like we'll still need human in the loop to validate the right subtitle results the right 80 script results and all that but still if we sort of divide that effort of having to understand the content versus plus the uh extracting that context into the output if that was what human was performing at least you get rid of this understanding the content part out and making that shareable, because in the human world, that was never shareable, I think makes financially a lot of sense. So it typically does. We run a cost analysis and obviously Chloe would bring our clients a lot of value in terms of cost efficiencies.
19:16Have you deployed this super broadly or is there still certain projects where Chloe is much more ideally suited for? And if so, which ones would be more suited for it? Well, I mean, our main business line has been subtitling. So we've been actively deploying Chloe internally on our subtitling business based on the approvals we get from our clients, their consents. Consent is, by the way, very important in our play. And we'll get to that later on. But anyways, so subtitling is the skill that, and by the way, subtitling is not just one skill because there's all sorts of languages. So there's like all these source and target language variations that we need to validate and, and, and, and roll out and improve sort of that, you know, feedback loop and continuous improving and all the AI world, you know, concepts.
20:11But anyways, self-titling is a skill that we're mostly working on. We've also deployed in a few clients that are, you know, using Chloe to, to generate dialogue list, spotting list, they call it, i.e. sort of the scripts they would distribute to vendors like us to do dubbing or subtitling, typically dubbing. So they're also looking at using Chloe internally to make that repetitive or fully manual task sort of a little bit more automated, but then with the right human in the loop. Yeah, so hopefully it's rolling out pretty rapidly, both internally and externally, I think getting the right traction as well.
20:56You did mention the permission, the consent, and yeah, I want to talk to you about that development with, I don't know, how would you say that? SAG-AFTRA, S-A-G-AFTRA, or how do people pronounce that? SAG-AFTRA? I think people call it SAG-AFTRA, yeah. SAG-AFTRA, all right. So there was this AI and voice actor dubbing development, right? So, you had the, they struck a deal around like explicit consent. So, I guess if I'm summarizing this correctly, studios need like an actor's permission to adopt the voice into another language using AI, you have these residuals. So, if AI translates an actor's performance actor, the actor, the voice actor gets some kind of required minimum pay, then there's no loopholes and training and synthetic voice generation limits, et cetera.
21:45Yeah, what are your thoughts on that? It's a very interesting discussion and I guess it's actually like a trailblazing, kind of pioneering discussion if you look at the much broader kind of macro AI picture. For us, it's really straightforward, to be honest. You know, these voice talents or actors have been the core part of, I mean, sort of an extended arm of ourselves, right? I mean, they have provided us and helped us as well as our clients to localize contents and not just to protect that traditional relationship but legally a performance voice is is their identity and their ip that we need to protect and and we live on that belief um we treat it that way and you know at the end of the day we need to find a way how in the new world where well we're sort of segwaying into ai dove which i don't think we're at that inflection point Maybe we are sort of at that inflection point, but, you know, more for the lower tier content.
22:44So that's another topic in conversation. But I still don't believe the premium contents could be replaced with AI. I think AI is still trailing, you know, sort of human performance in terms of the naturalness, the emotionalness and all that. But when we think about like these sort of digital replicas of voices, the way we see it is really the three Cs. and consent being part of that is we'd like to work with our voiceouts if the use of that ai voice comes and it's within the scope of you know the level of quality of services that we could offer to our clients and it fits into you know to our offerings you know the way we see it it's it should be based on first of all their consent of using their voices for that specific content obviously them having control which is to refuse whether i mean obviously if you're a voice actor for your voice to be used in a politically biased show or religiously you know you know what i mean right and it might not be what you want right so you would need to have control and obviously they need to be paid um for that specific content whether whether it's um an upfront payment or it's on a sort of a running royalty basis so really that reflects how we currently work with our current set of voice actors for every release we have aors that specifically details out what you know windows or mediums could carry their voices for whether it's theatricals or for streaming and it's based on these type of contents um Um, so, so that, that really is, is where we believe we could play a role whereby, you know, we would like to, you know, work together where we could, because right now that's effectively what we're doing.
24:45We're, we're sort of moderating in between the clients and the voice cons in terms of the AORs. Obviously we do the recordings and directings and translation as well, but then we help them, you know, in terms of those, keeping those three C's in place. including compensation rights. So we'd like to keep that position. I think that is the right way to go. And the reality is one might think technology could just replace all these actors. I mean, we're in a lot of conversations and POCs and tests, both internally and externally. It reveals you need properly recorded, properly acted samples and voices to be used for the purpose of the quality expectations that these core clients need for dubbing their contents.
25:38And that's not equivalent to some of the more crowd-collected, you know, publicly collected data points and engines trained or models trained with those sort of data seemingly, you know, fall really behind those expectations, whether it's the sound quality, whether it's the performances that you actually need in order to use the models to generate a very creative output, they're just not there. So I think there is the natural need for these actors to be part of that. And rather than being replaced by some synthetic voice, I don't think that worlds exist just Technically, it's just very hard to get there and legally as well.
26:26And we'd like to be alongside these talent communities to be part of that movement towards that future state. Again, depending on where the demand really evolves, I don't necessarily think Ayuno is going to be moving front and trying to offer these to clients. I think we would like to follow the clients and when the demands come up, especially the fact that we service most of the more higher tier entertainment world rather than sort of the creator created, you know, YouTube user generated contents or so on. I think we'll sort of follow the trend, but then we'll react as we go. But internally, we do a lot of experiments, tests to be beyond the game and react when that's necessary.
27:12Especially with things like emotion, naturalness, and other sounds. I mean, in some of the platforms, it now allows you to micro-control and add emotions. You still need the samples. So funny enough, if you use a voice that was sampled with mediocre emotion or performance, and then you prompt it to say yell or just cry, it doesn't because it just doesn't have the sample and it just is very mediocre. That's amazing. Okay. Yeah. Okay. That's super interesting. So what you're saying is if the actor didn't actually do a proper, yeah. Okay. Now I get it. I'm a little slow here. So you need to have the actual sample of the person that's basically whose voice is being synthesized.
28:00If there's no sample for that particular emotion, the AI is struggling to create that. Yes and no, because at the end of the day, it's a model. So it would somehow find a way to express that, but it's just not natural enough. The other portion is, for example, this is a little bit off, but if we do a cross-language conversion on voices as well, right? We've sampled a Korean version, and then we could generate an English version from that voice. We could. i mean there's a lot of models that could do that but then it's just awkward and the pronunciation is off because the way we use our our tongue and lips in in english may be different than korean so it just doesn't travel that way um still i think the models have to improve and they will improve to be honest um but i think we're still off from from being able to just synthesize a voice and that would replace all the voices.
28:57No, we're not there. We need good samples from possibly the professional acting community, whether it's voice announcer or other professional actors. And I think that fits in the interests of our clients as well. I haven't seen any client that's just trying to save a penny to just use any voice. I think the struggle is more towards the lower tier contents where they really don't have a lot of budget to spare. but once we build the right ethical model out that reflects these three c's and maybe our voice talents where actors can decide okay for this theatrical release i'm happy getting a thousand bucks but for this youtube i'm happy getting a dollar and maybe they support that as long as they can consent and control and be compensated i think it works it's just somebody has to build well first of all take that responsibility and that role and build the right platform to run it, which we are sort of internally progressing on those ideas.
30:00So I think those days will come pretty soon. Yeah, you're right. I had a similar experience once. I think I used an English recording of my voice and then AI dubbed it into German and it somehow gave me a Bavarian accent, like a really odd kind of, not even like every word was Bavarian, it was just like randomly every second sentence I had like this Bavarian accent, which was super weird to me, right? So the only data it had for me was my English voice, but then I had a different one in German, which, yeah, it was weird. And yeah, I guess what some people or many outside observers would underestimate is in the grand scheme of a production, of a kind of a top-tier production, yeah, you don't want to nickel and dime it by using some of these kind of advanced technologies if you want to have that high-quality output.
30:49All right. Well, let's go to less high quality output. There's been this boom in China in these AI generated micro dramas. And I haven't consumed too many of them because I don't know, since I left China, I don't have any active WeChat account anymore. But like, yeah, what are your thoughts on that? Is there any opportunity for like, you know, kind of premium dubbing companies if this catches on or maybe doing business in China from things like that? Yeah, it's funny because these microdramas, they initially started out with very lean production setups, lean budgets. But with the evolution of video, gen video models, they're really rapidly shifting to generating using AI.
31:37It's booming, yes, for sure. And it seems like they're very profitable as well. The funny thing is they're creating their contents with AI, but they want human dubs on this. We have a few clients. And our AI dubbing iterations never traveled with them. So they're creating contents with AI, but they need human dubbing. So that is reality. But then their budgets are pretty constrained. So we somehow need to find a cheaper way to support those human dubs. But it comes with implication, though, right? When we talk about all these AI doves, why can't they use AI doves on AI-generated contents? And if you ask me, it's pretty stating the obvious when they had to pick from all these variations.
32:27If you sort of imagine the Gen video process, very similar to the Gen AI dove process that I mentioned to you. It's very manual, by the way. So they have to prompt the scenes and they have to choose the best creative outputs, the best performance, whether video or audio, to make that content, right? That micro drama. And they've gotten that final version. Now to use AI dub, that would sort of, because it's all sorts of, there's no way they could generate the same version. and what I mentioned during SlaterCon, is every time you click a button, AI generates different results. You can't really creatively control the outcomes.
33:04It just has to be a repetitive task. And for AI-Dub to exactly reflect that performance in their creative intention is pretty much the same constraint as well-made contents in using AI-Dub to try and mimic their creative intent. It's just maybe that's why they just still need human for that. Who would be their primary audience? Would it be mostly kind of Southeast Asia or Korea or Japan? Blackbird Congress? Yeah. They're doing very well globally. I think they're doing 20 languages. US is a major market as well. I think I need to step up my content consumption game here. Appetite? I once read that I think initially that there was a lot of money to be made, but as people kind of caught on to it, I think it got super competitive.
33:53So I think the returns diminished a bit. I understand there's a lot of competition there as well amongst themselves, but the consumption is there. The viewers are there. It's just more competition on their side in terms of suppliers. It's my understanding. As he alluded at the beginning of this podcast, I mean, I, I, you know, grew significantly through a number of acquisitions. You mentioned BTI, you mentioned SDI. Do you still see opportunities for consolidation in the space or in kind of the narrower localization space, or would you be looking at kind of adjacent and acquisitions in technology or maybe other kind of adjacent production services?
34:28Yeah. Well, the truth is there's been an oversupply, especially during the peak we call 2022 peak, where there was just a lot of volume to localize, both in terms of dubbing, I guess more so with dubbing and a lot of the companies and competitors, including ourselves sort of over invested and over time that volume did normalize and truth is we still have a lot of residual capacity so in terms of capacity we're not so much you know keen on on on expanding more um through for in forms of acquisition in terms of coverage i mean in terms of language and global coverage you know we our recent acquisition was in turkey and brazil which pretty much fills all the gaps that we need.
35:21Obviously, when we didn't have these two operations, our clients oftentimes came up and said, hey, do you guys have anything in Turkey? Do you guys have anything in Brazil? And again, a lot of that is demand-driven, which is why we've closed those deals. But looking forward, I don't think we're going to be so much interested in expanding coverage nor capacity. Rather, again, we look at expanding ourselves towards supporting some of the lower tier or wider tier of work that we originally hadn't been doing, especially with the power of Chloe now that we could offer more sort of, we call it SLAs, but different tiers of service, all the way starting from sort of what we call Chloe Zero Shot.
36:04So because it has the contextual understanding, just upload a video, you'll still get a dubbing file or a subtabling file. That would be more for very, very low tier customers. I guess, but the fact that we could offer a more wider range of offerings, definitely we'd be targeting at sort of the wider part of our TAM. And then it does open doors for some adjacent markets, such as like audiobooks, e-learnings, or games, where traditionally, again, our studios have been sort of over-specced. it's just too high spec a studio to support their budgets but that opens up with Chloe and AI Dove for us so rather we're utilizing AI towards addressing our TAM towards the lower and adjacent sides of our markets yeah that is our focus for the upcoming year too.
37:08Is the primary channel for that mostly YouTube or what other kind of distribution channels do you see for that type of content? I mean, we definitely like to target YouTube, but it's highly competitive. There is what, like 100 different zero shot self-pending or dumbing solutions out there. So I don't necessarily think we want to compete with them. Obviously, I think we are the best when it comes to zero shot as well, but we'll just have to see for ourselves. I think our interest is more towards, again, human in the loop, creative in the loop, sort of leveraging AI to make it budgetary, more affordable so that we could address these traditionally clients that we couldn't access.
37:49But with human in the loop, so that it's still a professional service and it's sustainable and it's meaningful volume for the business in terms of a managed service of a standpoint. So yeah, a little bit higher tier, I guess, than YouTube's. And Microdrama is obviously a big target as well. So what's next for IUNU? 2027, what are some of the plans? You obviously spoke a lot about Chloe, but anything else? Or is Chloe the primary focus for next year? Yeah, I mean, obviously, Chloe's going to take some time. I mean, internal transformation is one thing. And, you know, any localization, media localization companies should know, but the client specs and workflows and everything's different.
38:39There's nothing called standardization in this industry. So to roll that out internally is a heavy lift. So that's going to take some time. I think we're doing it in good speed. to true but fundamentally what what where we'd like chloe to be isn't just you know making our subtitling operations efficient like i said it's really to leverage that contextual memory for for either clients and us to create multiple outputs so one source multi-use right and to get there there's a lot of i mean clients do see the value they're not gonna to tap into the whole picture from day one they're probably going to start off with the dialogue list spawning list okay i you know could use the contextual memory to generate subtitles then they'll get into sdh files and etc etc and then they'll expand as they go um just really getting that whether it's chloe or whether it's a different product getting that sort of future state operating model out there and sort of everybody accepting that this is the future I think it's at least going to take two years.
39:47So that really is where our focus is going to be. At the same time, obviously we'll be looking at adjacent markets to grow the business. But again, along that side to establish the right new sort of formation between AI and creative talents and human, both internally and externally, our employees as well as our creative talent networks, is really going to be the key in that play. Because, you know, again, I still don't believe there's a world where human gets totally out of the loop, especially for commercially offered media services, be it the high tier services or low tier services. So really establishing the right sort of new model.
40:35And obviously it comes with a lot of complication sensitivities and maybe some conflicts along the way. but taking that leadership and designing the right future for everyone, everybody so that they could adapt to the new world. Obviously, roles are going to change. Some jobs are going to disappear. But I think there is an opportunity for everyone to be part of the new model as long as they adapt. So defining that would be the biggest goal for me personally as well. Just before you close, you mentioned this now a couple of times that you do see that kind of human in the loop in the foreseeable future, which is a very different message than like I guess 75 % of the rest of the localization core or adjacent industry.
41:21Is that because in media, the localization is more part of the actual product in a sense, it's more part of the actual art of producing the final output? it? A voice actor is basically kind of actor adjacent. And then you could argue, well, why don't we just replace the actor? It's all AI, it's all AI micro dramas or macro dramas. Yeah, especially for dubbing, the models just aren't there, first of all. Will the models ever come? I think it's a different story. I mean, that's another long conversation, but it's just a sure amount of data that's just not there. You think about LLMs, they've effectively trained with all text data out there whether those are books or web-based data but they were readily available in pretty good format if you if you scan a hundred year old book and ocr that it's still going to be good information yeah but then if you scan like a like just a recording in a department store of people speaking it's just going to be very very bad quality data for it you know i mean in terms of multimodal.
42:29So just the sheer amount of data out there to train a good model has always been the constraint when we think about audio models or video models. But even video models are different. There are sort of good source of data that I guess the Chinese models have collected their data to train their own models. But when you think about audio, sheer good quality, studio quality data, where do we get that from? So audio is a pretty tough spot when it comes to the model evolution, if you will. You could get something that works, but when we think about studio quality, good performance data, and not just regular people, it's just, you could think that you may get it from the platforms or the video streaming services, but they're all mixed up with music and effects.
43:19And although there are technologies to demux that remove the music and all, it does get degraded. so um still i mean there this is one side of the story and just technically speaking and then there's the creative side of the story which floren you sort of mentioned but you know it's it's it's we live in a world we we can never say never um because things are so evolving so fast but up until now i i just can't be and i think the right balance is what what i see and and here for the near future isn't so much heading towards a full automation. I think that's going to take some time to get there. Great closer.
43:59Thank you so much, David, for taking the time today. Okay, thank you. That was fun. Thank you.
David Lee, CEO of media localization giant Iyuno, joins SlatorPod to discuss the company’s strategy, its contextual AI platform CLOE, and why human expertise remains central to media localization.
Lee explains how Iyuno is rethinking its operating model to deliver faster, more cost-effective localization services while maintaining quality. Rather than pursuing full automation, the company aims to integrate AI across its workflows while retaining human oversight and creative input.
Central to Iyuno’s strategy is CLOE, which uses frontier models and specialized agents to analyze audiovisual content and build a shared contextual memory of characters, storylines, and events. This enables the same contextual understanding to be reused across subtitling, dubbing, audio description, and other workflows.
Turning to AI dubbing, Lee argues that synthetic voices still struggle to match the naturalness and emotional depth of professional performances, particularly for premium content. He emphasizes the importance of consent, control, and compensation when using AI-generated voice replicas.
Lee also highlights opportunities in AI-generated microdramas, where some producers actually favor human dubbing despite constrained budgets.
Looking ahead, Lee sees opportunities to expand into lower-budget content and adjacent markets, including audiobooks, e-learning, and gaming. Scaling CLOE and establishing sustainable workflows that combine AI with human creativity remain key priorities.







