#268 Thordur Arnason on Why Capgemini Is Building an AI Speech Translator

0:00Thordur Arnason:Predominantly, Capgemini runs on a consulting model, meaning that we sell time for money, basically. So we normally don't build products that much. We normally don't do big ventures into other models. But this is shifting gradually now that AI is turning most things upside down.

0:25All right.

0:25SlatorPod:Welcome back to SlatorPod. But today on the podcast, we welcome Tordor Arneson. Tordor is the global AI go-to-market lead at Capgemini Invent. And Capgemini Invent is the digital transformation, digital innovation, and consulting arm of Capgemini. And Capgemini is a small French-based multinational tech services and consulting company with what Tordor just told me, over 400 ,000 employees globally. and Tortor recently launched an AI speech-to-speech translation product called BobbleSpeak and that really put them on our radar. So we wanted to know why a consulting company would launch a Language AI product and ask Tortor to tell us more on the podcast today.

1:08SlatorPod:So thanks so much for joining. Thank you for having me. So where in the world are you recording this from today? What part?

1:16Thordur Arnason:So this morning I'm sitting an hour north of Oslo in Norway at an old farm near the biggest lake in Norway with beautiful views and weather outside. So that is my safe space.

1:29SlatorPod:Safe space. And the sun has already risen at 9 a.m. or you're still waiting for it? No, we have sunlight. So that's good. All right. Yeah. Spent some time in Norway and I remember those dark mornings and early evenings. So in the winter. All right. So tell us a bit about your professional background in tech leadership and your professional journey so far with Capgemini.

1:54Thordur Arnason:Well, in brief, out of uni, I was trained in software engineering and started my career precisely there, doing coding for a living, which was fun until I got promoted into management after about two years. And ever since then, we're talking about the late 90s here, just to give you context. I've been mostly in management and executive roles in the consulting industry, but also some years in what you would call the more creative consulting setups like design and experience firms. Been with Capgemini for about eight years. I led the Nordic part of Frog, that is the design agency of Capgemini. But now, since 22, I've been focusing solely on AI, and that's why I'm the lead on go-to-market and AI.

2:52SlatorPod:Just on a bit of a tangent, Frog was a big client in my pre-Slater days for my previous company. So when did Frog become part of Capgemini? Is that a long time ago?

3:05Thordur Arnason:Well, you know, in these big corp things, it's a funny story. Sorry. So Capgemini acquired a consulting firm called Altran a few years back that previously had acquired Frog. So Frog was kind of in the portfolio that got acquired by Capgemini.

3:25SlatorPod:Huh. Interesting. Just side note, they were working on Singapore Airlines kind of website redesign like in the early 2010s and we were part of that. So yeah, very impressive company, Very interesting. All right. So, but tell us just for the listeners who may not be familiar with kind of Keptgemini's just general business model, tell us a bit more about that. And then, yeah, we want to segue into the Baublespeak launch that you announced a couple of weeks ago.

3:52Thordur Arnason:Predominantly Keptgemini runs on a consulting model, meaning that we sell time for money, basically. So we normally don't build products that much. We normally don't do big ventures into other models, but this is shifting gradually now that AI is turning most things upside down.

4:15SlatorPod:That's what kind of turned my head. So I'm glad that I wasn't totally off, right? So yeah, you sell hours for money, but typically you don't build products. So what triggered you to launch a kind of standalone language AI product? and tell us more about the product in general.

4:34Thordur Arnason:A little more than two years ago, I also had a role in something called the AI Futures Lab in Capgemini, looking at emerging technologies within the AI field, new models, all that fun stuff going on. And we were doing some research on what we now call multimodality. modality. So voice as a modality that only two years ago was not that great. But at the time doing research we came across a new model from FAIR that so that is Mera's now canned research arm in AI, they released a model called Seamless M4T. And Seamless M4T showed us that doing near real-time translation with voice as a modality would be possible in a near future.

5:32Thordur Arnason:So that started it. And me being an old sci-fi fan immediately thought about the beautiful concept of the Babelfish from Hitchhiker's Guide to the Galaxy. If you're not familiar with it, it's basically a small fish you put in your air, and once it's in your air, you will understand not every language spoken on Earth, but in the universe. So we took that as an inspiration, and that's why the solution is called Babel Speak, a little homage to Douglas Adams there. And we started building based on the work that FAIR had done and also combining other models and engineering we had done into an early proof of concept of near real-time voice-to-voice translation.

6:24SlatorPod:So what's kind of the motivation for this? Is it more of a kind of a tactical show of product capability? Is it a bit of a long-term vision for Capgemini? Is it something you're looking to monetize aggressively? Well, it started out in the lab.

6:40Thordur Arnason:And the lab has primarily not a commercial focus, but it has an exploration focus. So in the labs we run around the world, they explore new things. And sometimes we come across models or systems or technology that shows promise. And we then lift them out of the lab and start thinking about how we can potentially commercialize this. and that is the story of BabelSpeak as well.

7:09SlatorPod:And you chose like speech translation specifically because of that fair model. So that was the story, right? You came across this, you thought, well, what are we going to do with this? And then you chose that particular use case, right? It wasn't like you had the use case in mind and then looked for ways to solve it.

7:27Thordur Arnason:No, so back two years ago, we were doing a lot of experimentation on translation, but with text as modality. So, you know, the normal LLM kind of things you do. And even back then, though it seems like forever in this space, we were doing pretty accurate translation across dozens, 50 plus languages that were fully usable to be used in client projects. So LLMs and text as a modality seemed like not a topic for the lab to explore. So we shifted to the hard one, which was voice at the time. So that was the reason for us trying to solve something that wasn't solved before.

8:14SlatorPod:And you chose, in essence, like a horizontal use case. It's quite broad, rather than maybe like a domain-specific AI product. Why is that?

8:24Thordur Arnason:Language in itself and the abilities of degenerative AI models on manipulating working with languages is per definition a horizontal. This is something that is relevant everywhere in every industry, in every aspect of both professional but life itself. So in its nature, it's a vertical thing to think about. So we quickly discovered that the unmet need of being understood is tremendous. So we have translation services, we have professional services, You have your Google Translates and all that stuff. But still, there are big gaps in providing reliable and safe translation services across a bunch of industries.

9:14SlatorPod:And so when we did the research, I think you already ran some pilot projects with organizations like the DNB Bank, Norwegian Red Cross, Norwegian Police. Tell us a bit more about those initial pilots, what they involved, outcomes.

9:28Thordur Arnason:Yeah, so once we took our research out of the lab, we went looking for a client to do a proof of concept with them on a real-world scenario. And at the time, I had discussions with DNB, largest bank in Norway, on a very specific use case. Back two years ago, we got a lot of Ukrainian refugees coming to Norway, as we all did in Europe. And the bank had a very specific problem. Quite a few of these refugees, they needed banking, but they didn't speak, for sure, didn't speak Norwegian, and many of them didn't speak English. And none of the bank employees understood Ukrainian. So you had a group of people, refugees, with an acute need of being integrated into society and we couldn't understand each other.

10:21Thordur Arnason:And of course with the influx of refugees from a specific area, the capacity of translation was absolutely non-existent. It was overbooked like months and almost half a year in advance. So we didn't have enough translators to solve the actual problem on the ground. So we shaped a very distinct use case for the bank, making bank employees and Ukrainian-speaking potential customers understand each other, voice-to-voice. That's it. And we built that.

10:58SlatorPod:And the platform was just an iPhone or a mobile?

11:00Thordur Arnason:The platform was both on mobile, but also on your laptop. So the rep from the bank would typically then either put his or her phone on the table or use his or her laptop. And we ran this as a web app to make it flexible.

11:19SlatorPod:And I guess the Red Cross and Norwegian police would have been the same case?

11:25Thordur Arnason:So this, with the DMV, that was a very narrow case. With the Red Cross on intention, we went to them and started to collaborate with them on this half a year later, because they had multiple needs in multiple domains and situations and contexts. It's a way more complex set of use cases. And we wanted to work with the Red Cross because they are heavily exposed to situations where translation services are needed ad hoc, where you can't wait, there are acute situations, and that was a good fit for this technology. So we've been working with them since the beginning of 25, so almost a year now, and are going to run a pilot with them towards next summer because it takes time to learn in this space.

12:15Thordur Arnason:The police, about the same thing, a little less complicated because we had distinct use cases, border control, in jails, stuff like that where we could shape the use cases a bit more. But same story, needed to expose the technology to real world use cases.

12:36SlatorPod:Got it. And you also said, when we spoke to you for our coverage on our website, you said that it was developed in partnership with NVIDIA and Telenor. So tell us more, like, so NVIDIA for obviously kind of the underlying tech and Telenor as the channel or? I need to do a bit of explaining. Please do. Yes.

12:57Thordur Arnason:Yes, so we will never be able to compete, nor do we want to compete with services like Google Translate. Obviously, they and their competition have that space. What we shaped BabelSpeak to be is something that runs in as secure an environment as possible. meaning that we needed to find a place to run the service from that was sovereign and secure and that's where Telenor came into the picture because at the time we were also helping them build the first so-called sovereign AI factory in the Nordics and that was the perfect place to run Babel Speak because doing that we could guarantee, like the police, like the Red Cross, like a lot of entities in the public sector, other sensitive industries, we could guarantee that we own the pipeline, we have full control, it's sovereign on Norwegian soil, you don't have to worry about GDPR, you don't have to worry about security because it's running in a critical infrastructure data center.

14:09Thordur Arnason:And all these normal worries you have when you want to use services like this in sensitive situations, we took away by putting it there.

14:21SlatorPod:That makes a lot of sense. I was wrong with the channel. So it was basically, yeah, I mean, the hosting, the actually underlying... Yes. Got it. Got it.

14:29Thordur Arnason:And NVIDIA, obviously, are other GPUs that this runs on. in the AI factory, but we're also working with NVIDIA on the software stack side. So we are exploring their models because they also have translation models. The Riva suite is pretty good. So we are continuously exploring together with NVIDIA how we can use parts of their stack in our service.

14:57SlatorPod:All right, I want to get back to Bobble Speaker, but I want to run something by you that I saw on Twitter slash X literally a couple of days ago. and I want to get your thoughts on it. So there was a post and I hope it was real, but one of the major consulting firms said they advised an investor on due diligence for an AI startup. And the consulting firm says they were able to build a prototype of that startup's product within like two weeks that was better than the original product. So naturally the client of the consulting firm, the investor, didn't end up investing. Is this something that has ever crossed your desk?

15:33SlatorPod:where you're looking at something as part of a consulting engagement, say, well, that's quite easy to do. We could do this in like a week or two. Yeah.

15:42Thordur Arnason:This is what we're in the middle of right now. This is quite dramatic. And no, it's not just someone bragging on X. It is a real thing. So listen, I'm a senior, quite old VP, Jude, with engineering degree from the 90s. I am building products as a side gig to show to my clients now. I'm doing stuff in five days that would, two years ago, would take me and a team three months to do. This is what's happening right now. And that is both very cool and very scary at the same time.

16:24SlatorPod:Where's the moat then for something like kind of a B2B or B2 government speech translation product? Where do you think is a moat? I mean, you mentioned the sovereign aspect. Maybe there's, obviously there's some quality issues, language coverage issues, but do you see any other moats that would be like long-term defensible there?

16:45Thordur Arnason:Long-term, I don't know, but at least for now, our moat is all the engineering effort we put into this pipeline, making it real-time, because that wasn't easy. It wasn't easy back then. It wasn't easy this year either. So we have been able to work down the latency. So the time it takes from me to speak to you to hear from three seconds to less than 300 milliseconds by hard engineering work together with NVIDIA. So optimizing a product like that takes a lot of effort. And the changes we are seeing now in tooling and what people can and cannot do does not affect that because this is hardcore engineering.

17:30Thordur Arnason:That's one moat. The second one is the sovereign play, obviously.

17:34SlatorPod:And when you go with the sovereign play, like are these, let's talk about maybe Europe, are the Telenors, the Swisscoms, the Deutsche Telekoms, would those be kind of the preferred partners for something like this in Europe?