#292 Voice AI Data for Lower-Resource Languages with NCSpeech

0:00Iurii Agafonov:We saw the opportunity, we saw the market gap in voice technologies, in data collection, especially in emerging markets.

0:12Dmitrii Sandzhiev:Hey everyone, welcome to another very exciting episode here on SlaterPod. We are on the podcast today with Dmitrii Sandzhiev and Yuri Agafonov. So Dmitrii and Yuri are the two co-founders or co-founders of SpeechAI startup NC Speech. We love to have early stage startups on the podcast to discuss how they see the industry and the world and where things are heading. It's always good to check in with early stage startup founders. So, Dimitri and Yuri, thanks so much for joining today.

0:41Iurii Agafonov:Thanks. Thank you very much for having us.

0:44Dmitrii Sandzhiev:So, maybe the first question that I always ask, like, where are you recording this from? What part of the world, what city, what country?

0:50Iurii Agafonov:Currently, I'm in Kuala Lumpur in Malaysia. It's Southeast Asia, quite far from your location.

1:00Dmitrii Sandzhiev:I am in Kazakhstan in Almaty. It's a city named the south capital of Kazakhstan. That's two very interesting locations. I mean, KL, I've been there a couple of times when I lived in Singapore. So they still don't have a good train going up there, I just found out. But you can take the plane and the bus. So it's one of the most obvious train lines to be built on earth, but somehow they don't manage to do it.

1:25Iurii Agafonov:They still have some plans. Every day, every year they announce that we're going to launch this train again.

1:32Dmitrii Sandzhiev:Oh man. And I saw that Singapore built this kind of green park. They gave up. They had the rail line all across Singapore forever. And then now they've just given up and they kind of made it into a city park. Anyway, different topic. Yeah, Almaty never been, but definitely one of the cities that you should meet should go. Cool. All right. So, Kazakhstan and Malaysia. So, Dimitri, you've built a lot of different things in your career. Maybe catch us up on kind of how you ended up or started now with NC Speech.

2:09Iurii Agafonov:Yeah, thank you very much. Well, first of all, I truly believe that my biggest achievements are still ahead of me and I have a strong feeling that one of them will be connected to NC Speech. And actually, yes, after graduating, I have a background in economics and I spent years in investment deals. And at some point I moved to another part of the and started my experiments as a founder. And I also strongly believe that being a founder or being an entrepreneur, everyone should have a skill of feeling the opportunity. And actually this is the exact feeling I had when I met the team. Actually in our team we have like today we have like around 12 people and six of them are our co-founders.

3:07Iurii Agafonov:We are all co-founders, so me and Yuri, we are just two out of six co-founders in attendance speech. So when I met this super deep tech engineering team and we saw the opportunity, we saw the market gap in voice technologies, in data collection, especially in emerging markets. So that's how we started in 2025.

3:33Dmitrii Sandzhiev:Got it. Yuri, give us the rundown of your experience with research and also, I think, as a data scientist. I was studying math and IT since I was a student. And that path leads me to some research lab, my first place of work. And it was voice research lab. So we did a lot of research there connected to automatic speech recognition, speech synthesis, voice biometric technologies. And also, me personally, I was focused on innovative things always. So we did modern acoustic search algorithms. I personally developed it. and it was actually my master thesis. Then we tried to make speech analytics system which helps people to analyze what's happening in their call center, for example, or on a gas station.

4:40Dmitrii Sandzhiev:So yeah, basically that was the start. When we had, with our team, we had an opportunity to join a startup, which was real-time voice conversion technology. Actually, not voice, but whisper-to-voice conversion technology. So that's much more difficult task. And we successfully solved it. So, yeah. So basically, I was working on optimizing neural network inference on devices, different devices, Windows, Mac, phones, training neural nets, like automatic speech recognition, parametric nets, voice cloning nets, all that stuff. Yeah, so I have experience in OLM training, text classification system training, so quite wide experience.

5:37Dmitrii Sandzhiev:And now, Tennessee Speech, I'm continuing what I was started from. Cool. So let's start with the kind of the elevator pitch for you guys. So give us the elevator pitch for NC Speech and yeah, maybe, I don't know, 30 seconds, 90 seconds, if there were VCs listening, how would you describe it?

6:03Iurii Agafonov:Depends on the elevator, right? Where we're stuck.

6:07Dmitrii Sandzhiev:In KL, let's say, I don't know, KL Central, like so maybe a hundred stories. We go up, we have a minute.

6:13Iurii Agafonov:At AnsysPitch, we turn idle time in super apps into AI training data. So we allow drivers, riders or passengers to monetize the idle time to get a reward by submitting some data. Let's say it could be audio data, photo data and video data. These datasets are sold to AI labs, to enterprise R &D departments and to sovereign AI initiatives. Today, Uber and DoorDash, they have started this AI digital task like this year. They have already proved the market. We became the plugin layer for everyone else. So we connect the demand from AI Labs and from other data customers with real-world environment data that can be collected through our partners.

7:08Got it.

7:10Dmitrii Sandzhiev:Well, now we are on the 101st floor and we exit and we look at the beautiful landscape in KL. So what kind of technical breakthroughs made this possible? This is something probably that five years ago would have been not something that would be top of mind to build. So in the past two, three, four years, what kind of specific breakthroughs made you think that this is a viable path for a startup?

7:35Iurii Agafonov:Starting from the ground, first of all, we have identified that today there is a huge demand for data, for data sets, for AI model training, especially in emerging markets, especially in Malaysia, right? So in Malaysia, only three main official languages, around 10 different dialects within any language. And when you build something in voice AI, so if you want to build like voice AI robot for call center automation, you need to have a reliable model that can easily at least understand what people say. Not like when they speak like in academic Malay or English or Mandarin and so on, but how they speak at the street.

8:16Iurii Agafonov:So, and when we faced this problem, we decided to move, to move like once, make one step back and start with a data collection and model training first. And then, then we also have managed to onboard a big client here in Malaysia who faced the same problem, right? So, and we started to collect and collect process and sell these data sets for improving like voice part of sovereign AI here in Malaysia. The same we do in Kazakhstan and with our partners today we are scaling these operations within emerging markets and assist AI labs, enterprise R &D, like sovereign AI initiatives, we assist them to collect and improve their models faster and more efficient way.

9:16Dmitrii Sandzhiev:I would add to Dimo's speech that in the last two three years we all of the world entire world saw the abilities that I can provide so the I mean this open AI success anthropic success, but it's mostly now in text field. And what is greatly uncovered is voice field because this data is much harder to collect, especially in different languages. So let's dwell on the language for a second. I mean, you mentioned Kazakhstan, Malaysia. So tell our viewers and listeners a bit more about the language set up in Malaysia and also in Kazakhstan. I mean, in Malaysia, you have various, I mean, you have the obviously English, Malay, probably Chinese you cover, but what are some of the particular challenges there?

10:16Iurii Agafonov:Yeah, I can give you an example. The business case that we have solved, there is a bank in Malaysia, a bank with a huge amount of customers. they have of course like bank like application and they wanted to integrate voice control mode to the application so they want to allow people their customers to control they like bank account transfers check balance using voice not like typing not like in a traditional way but using voice to solve this problem, it means that this application should understand what people say. So, and when you, I mean, definitely you can say like transfer 25 ringgit to my mom, you can say it like in very different way, in very different like conditions.

11:11Iurii Agafonov:You can, especially in Malaysia, you can switch languages, you can switch from Malay to English and then back, or if you are like Chinese ethnicity you like switch from Mandarin to English and back. So and the problem was when we started the one of the most important metrics in like speech technologies is like word error rate. So it means the percentage of when the model understand what you say. At the beginning this water array was around like 40-45 percent so meaning that in 45 percent model literally don't understand what people say and when you when you are like very serious like big enterprise as a bank you cannot rely on this one on this one right otherwise all your customers will will face uh many problems right you start to transfer to different people or do some like mistakes.

12:08Iurii Agafonov:After we collected around 300 hours of data, specifically collected for this bank, for the banking domain, for financial domain in Malaysia, collecting from different languages from more than 1 ,000 contributors, we used these data sets for improving this speech recognition model. So after After this project, the award rate went down from 45 % to approximately 15%. So 15 % means that you already can use it in a commercial way. And this is a very small example for one bank in Malaysia. But another bank in Malaysia is going to do the same, right? Another insurance company, telecom company. So we only look at very small, like 30 million countries, Malaysia.

13:08Iurii Agafonov:You can imagine what's happening in Indonesia, in Philippines, in Cambodia, Uzbekistan, Kazakhstan. So all emerging markets today, they face the same problems. They want to use this technology, they want to integrate it, they want to somehow use AI in their business processes. but they need to build some reliable and confident and safe solutions for them.

13:37Dmitrii Sandzhiev:What's the data infrastructure part here and what's the kind of maybe manual or somebody else does the manual collection of the data part? So what's the scalable part and what's kind of more the solution slash services part in this? About infrastructure, it starts from the device you record with, if we're talking about audio. And that has a huge impact on the data you're collecting. That's the first part. It's already, you have to control it somehow. then the environment, then delivering to the platform, which is if you're not in acoustic studio, when it's remote. So like here in our podcast, as we mentioned, there is some nuances, let's say.

14:34Dmitrii Sandzhiev:Let's stop here. So, yeah, that's the part of services, I think. So how you develop that, how you upload that, how you process that later. So that's here is the key part because you have to control the speaker stability between records. For example, some users may ask his brother or his mom or some other relative to Oh, continue. I don't have time, but we have to control that. there is one speaker because we have in our data, which is, uh, it's very important to control it. Or maybe when they try to, uh, use some speech synthesis algorithm to do the work for them, and that's not what we can accept also.

15:25Dmitrii Sandzhiev:So there is lots of services that working on, uh, data quality that we collect, but at the same time there is an infrastructure where you have to storage there somehow, you have to store it efficiently. For example, if you're collecting data in Malaysia, you don't want to store your data in the USA. At least this data will go across the half of the globe to reach the server and then signal goes back and you probably will have some issues with government which usually governments not like when data of their citizens stored in another country so there is lots of problems issues you have to solve and yeah that infrastructure is kind of part of it.

16:21Dmitrii Sandzhiev:And of course, there is infrastructure for processing this data. So you need GPU, you need probably some cluster. It depends on the volume you process. Yeah, so basically I would say that it's connection infrastructure, storage infrastructure, and processing infrastructure. And how you organize this internal. I wanted to add something. Like you mentioned that, hang on. So speech synthesis. So some people in, I guess, your crowd or the people that are delivering the data, what, they're trying to submit data to you by using, I don't know, copy, paste, text, and then have speech synthesis. And then, what, you would get AI-generated data and they try to trick you that way?

17:09Dmitrii Sandzhiev:Or, like, how does that work? Yeah, that's a possible outcome. And we have to control that because it's a threat to data quality. It's so funny. Like even in an early stage startup, like you get these kind of probably edge cases like immediately. Like you'd think like, oh, people will contribute the data. And then no, no, no, they're uploading AI generated data. It's like, OK.

17:32Iurii Agafonov:There is a lot of cases of abuse, of abusing the system because it's on a reward base, right? So when you give this opportunity to the crowd, definitely a lot of people started to abuse your system to get the reward. And what Yuri just mentioned is we have built this kind of technology pipeline for quality control. So we manage, we can control the quality, we can manage this quality and we are able to convert raw data to like training ready data sets. But this is only one part. It's like this is technology part. Another part is how you are going to hire and onboard thousands and thousands relevant people, even on the reward base, right?

18:23Iurii Agafonov:you need you need to spend like money you need to spend time on marketing on on this hiring so you mean it means you have your customer acquisition cost even just to deliver the message the information that guys you can get uh some revenue by like submitting data so and this uh is our i think this is like one one of our mode and uh what we have like elaborated we came to right failing and delivery super apps so the companies they have absolutely different from like ai from ibis they have like core business delivery like taxi driving and so on right but what they have i mean the most important at least for us is uh real people who is every day in real environments and this is a kind of geek platform like geek platform economy so people come to be become a driver or courier or rider anyone they come to get some uh to as a profit like to get a revenue to get a revenue for for them and when we came to this super apps platforms we came with one very simple uh value we can give you the extra revenue stream for your drivers for your riders when they have time like idle time during a day and so on so uh and it was i think very like good product market fit because right now we have we already have like very one active uh partnership with right handling and delivery application it's a global one it's already presented in 48 countries so definitely technically we have access to all these 48 countries to collect data.

20:13Iurii Agafonov:We started another collaboration in Southeast Asia very soon with another Super App. But idea is so we bring some value so drivers get a reward, we get the data and we also share the revenue with our partner with the super app. So, and we, by this approach, we have solved this problem with hiring and onboarding. And we have zero customer acquisition cost. Because if you already are using your right-handing application or delivery app, you already have your application, right? So when you, let's say, order a taxi and you go somewhere, you have kind of captive time, you have 20 minutes of your time when you just sit and do nothing.

21:02Iurii Agafonov:You can definitely can scroll some social media and so on but if at this moment we can send you like push notification like, Florian, do you want like to get 10 % discount for this ride? You just need to spend like five minutes of your time by doing something like doing a very simple micro task. Well then it's like up to you, right? If you want you go and proceed. If you don't want, I mean it's opt-in, right in opportunity. And looks like that we have found this very good match how we can collect it and that's why we are even now we are as a like small startup we're already quite fast ready to scale like super fast with this like with this capabilities of our of our partners.

21:53Dmitrii Sandzhiev:Let's assume just to make this clear to the audience. So I'm in KL, I'm ordering a grab, I step into the grab, I and then the driver would kind of prompt me, hey, would you like to do that? Or is it like, I know I get it. So okay, then not the driver. But if I have push notifications for grab enabled, it will pop up, it will give me the option. And then yeah, and then I get a discount on the grab For example, yeah.

22:22Iurii Agafonov:For example. Or some loyalty bonuses that you can spend in Grab ecosystem. And Grab's incentive to do that is what? For most of Geek platforms companies, the main challenge they have is to provide better conditions, better opportunity for their riders, for their couriers, for their customers, first of all. And if they provide you, let's say, you mentioned this Grab, right? If they provide you this opportunity to get like some extra, something extra, right? Like loyalty bonuses or even like real money, real ringgits. I mean, it depends on the structure. You will be more thankful and more loyal to Grab.

23:16Iurii Agafonov:So next time, if you get loyalty bonus, next time when you need to order a taxi, I mean, what would you choose, right? To choose Uber or choose maybe Grab because you already have some bonuses that you can spend within Grab ecosystem. So we're increasing the loyalty of our customers, of SuperApps to this SuperApp. Because for you, from a client perspective, it's coming from SuperApp, not from NC speech, it's coming from Super App.

23:47Dmitrii Sandzhiev:All right, very, very, very interesting. So we're talking about speech mostly, or is, for example, for the example we just went through with Grab, like is it mostly speech or is there other things they would be doing there like image or like other kind of data sets that you process? Or do I literally like think about like somebody in the backseat of a Grab, like kind of processing

24:10Iurii Agafonov:speech data from technical point of view we can collect and process different modalities yes of course speech is our like core expertise like yes but we also can everything you can do with your smartphone using your camera using your like mic microphone of your smartphone uh geoposition so everything you can do with your smartphone this data we can collect so another example for okay for speech technologies is quite clear right we collect like scripted speech we can collect dialogues we can collect spontaneous speech for like companies like who is building like maps for example uh they have a request for like confirm uh for collecting uh photos of like some point of interest like take a photo of like some uh door for to the shopping mall or take a photo of some like specific object from today, from current perspective.

25:15Iurii Agafonov:And we can give this opportunity to customers to take a photo, what is needed, and then also collect the rewards from it.