Latest / Convo AI World / The Startup Using Voice to Detect Depression, Fatigue and Diabetes

Transcript

The Startup Using Voice to Detect Depression, Fatigue and Diabetes · Convo AI World

• Emilia Molimpakis (00:00) It used to be 10 minutes. Now it's down to 15 seconds and it's real-time processing, so you get your scores immediately, basically. We embed our models within vehicles to detect driver fatigue and stress in real time to help avoid accidents. That's super big. This is a huge opportunity for us to provide essentially a health and safety infrastructure layer for all of those voice communications. So we are enabling safe voice agent deployment at scale If you told me six years ago, you could hear diabetes through somebody's voice, I'd be like, "Are you crazy? That's not possible". It turns out it is. Hermes Frangoudis (00:41) Hey everyone, and welcome to the Convo AI World Podcast where we interview the founders and pioneers pushing the voice AI space forward. Today, I'm very excited for a special guest, Emilia from Thymia. Welcome Emilia. Emilia Molimpakis (00:56) Hi everyone. Thank you so much for having me. It's lovely to be here. Very excited. Hermes Frangoudis (01:00) To kind of get the ball rolling and get things started, I'd love to hear really what inspired Thymia and like what inspired you to start this and get involved in this space. Emilia Molimpakis (01:11) Yeah. So I guess a quick, just short intro on what Thymia actually does. So Thymia is a deep-tech company. We're based out of London. And we specialize in using voice as a biomarker for your health with applications kind of very far and wide across healthcare through to safety-critical and so on. What actually inspired the founding of Thymia. So I founded Thymia about six years ago now. Prior to that, I was a researcher for about 12 years. Absolutely loved my research. I specialized in using language as a biomarker for cognitive function. I was a neuroscientist at UCL for a long time. And while I was doing my research, my best friend, who is also an academic, actually developed depression. And I saw her firsthand kind of go through the whole process in the UK of trying to get help. She went to the NHS. She saw a GP. Eventually, she saw a psychiatrist as well, a very good one. But unfortunately, even he didn't realize how bad her condition was. And just two days later, she ended up trying to take her own life. And that day we were meant to be meeting up. She didn't show up and I went to her house and I was the one who found her. And as you can imagine, that is very much a life-changing experience. Thankfully, it ended up being positive. She's doing a lot better now. But it completely changed how I thought about everything I was doing. It made it very clear that if we don't take all these amazing things we're finding out in research and actually give them to people who can use them in clinical and kind of the real world, then there's no point in what we're doing. And so I instantly quit my postdoc. I think it was maybe within a week and decided to try and help other people since I couldn't help my friend. And the idea was to basically use voice as a more accurate, objective biomarker. Initially, it was for mental health, starting with depression, and then it kind of expanded from that. I didn't have all the skills myself to build Thymia from scratch, so I joined an accelerator program, Entrepreneurs First in the UK. And met my co-founder there, Stefano, who's the CTO with perfect complimentary skill set. And that was six years ago we started and yeah, that's it. Yeah. Hermes Frangoudis (03:18) It's an amazing story and the catalyst, while that must have been a very difficult time for everyone, it's kind of an amazing dot to connect, right? Like, how do you take something that you've been studying and learning about and really applying it to the real world? Can you tell us a little bit about that science that shaped Thymia's early hypothesis and kind of how you connected the dot, I guess? Emilia Molimpakis (03:42) Yeah, absolutely. So I guess there's a lot of history, kind of research-wise that links voice or speech in general to mental health. And that was kind of where we were starting. So speech, you can think of speech as being multimodal in itself. It has different aspects. Think about it with three layers. The first layer is kind of acoustics, the sound wave, the physical properties of speech, like how do you actually sound, loudness, friction, intonation, frequency, whether your voice is hoarse, how your muscles are actually controlling the vocal cords. There's literally thousands of features in the acoustics of your speech. Then content, what are you saying, content in terms of words, meaning, structures, and then timing. So pauses, breathing, like the slowness or the latency of your speech, each of those aspects is tied to different parts of not just your mental health but actually turns out, the entire nervous system. So for depression, for instance, when somebody is experiencing a depressive episode they may have a lot more flat intonation, they have low mood, they have fatigue and that is expressed through voice. There's a lot more personal pronouns that somebody might use because there's a lot of introspection so there's a lot of features there. That was basically what we used initially to go into building our dataset for mental health and building the original models. So starting with depression and anxiety, we knew what we were looking for. And then we just expanded, gathered a massive dataset to build this properly. And yeah, started from there and then went on to not just depression, anxiety, but many other aspects of mental health. ⁓ In growing our dataset and the capabilities of our models, we realized it isn't just mental health, it is the entire nervous system. So we can look at respiratory health, cardiovascular health, and even metabolic health. That's actually my favorite right now. We've started finding signal for diabetes type-2, which is incredible in itself. I can go into that if it's of interest. And yeah, the applications have just kind of skyrocketed from there. So based originally in my research and then expanding as we go. Hermes Frangoudis (05:45) No, that's amazing. So you were able to kind of like build that foundational like biomarker model based on data you've assembled and you knew what pieces to kind of look for in the haystack, right? Emilia Molimpakis (05:58) Exactly. It's, I think people really overestimate how easy it is to do this or underestimate how hard it is to get relevant health signal from speech, from voice data. I think, okay, you know, somebody like OpenAI or Anthropic, like they have access to, you know, they can just scrape the internet. They have access to so much voice. Surely they can do this. It's not that easy. You need very specific populations, very specific types of data to account for multiple various variables. Human curated data is extremely, extremely important. And so it was that knowledge from research of how do you gather this dataset? Who do you look at? How do you get them to do longitudinal data collection with you? This is people with depression who don't want to engage with anyone. How do you, how do you engage them? And realizing that voice alone isn't going to help us build a dataset. We need multiple types of data to fill in the gaps. So our dataset is actually multimodal beyond voice. It's video data, behavioral data, EEG data. Now it's blood biomarker data. We're going to be doing urine sample data soon as well. It's all of that together longitudinally from around the world to build this really strong dataset. And that's why people don't necessarily, they haven't heard of us until very recently, it's because it's taken a long time to build this and to get to where we are. Hermes Frangoudis (07:20) I could not even imagine just kind of like the hurdles you had to jump through and ways you had to engage to get this out, right? Because everyone I speak to on the podcast, really it does fundamentally come back down to the dataset and how you've pioneered within your own industry to collect just the right data and having that foundation from the academia side of it, like sounds like it was really valuable to your journey. Emilia Molimpakis (07:45) Definitely. And when we started, there were no platforms to enable us to gather this type of data. We had to build our own research platform from scratch to conduct our own trials and research. So, Stefano, the first year, it's just the two of us, he built the whole thing from scratch. We started with video games to gather data because that was kind of how we would get around people's reticence to give us voice and video data. And it worked. It was a big gamble. Hermes Frangoudis (08:02) Okay. Emilia Molimpakis (08:12) Spent all of our initial money just on developing these interactive gamified activities to get the data we needed and it really paid off. So yeah, it was huge learning curve, but yeah, the academic background was super helpful. Hermes Frangoudis (08:24) I want to go back to something you kind of touched upon and bringing in all this multimodal data, right? And then being able to get the model to understand how it brings all these kind of markers together, but then uses voice as the actual like input of the biomarker to connect the dots for this new individual. Can you tell us about like how important that that is and like what, what it can really tell us about our health? Emilia Molimpakis (08:47) Yeah, so voice, turns out, like those different modalities that I spoke about, content, acoustics and so on, it's extremely powerful. I mean, I knew this from my research, but I don't think we ever anticipated just how powerful it could become with the power of AI and the right dataset and so on. So voice is super powerful, so very strong. Like you can see multiple facets of somebody's state of body, state of mind. But the other big asset of voice is it is everywhere. It is ubiquitous. Everybody in the world, or rather, not everybody, but the majority of people in the world will use voice to communicate. So it's very natural. It's very comfortable. It's an obvious way of actually communicating. And that also means it's very easy to access. People naturally speak. They don't naturally necessarily turn a video on or anything but talking on the phone, et cetera, people are very comfortable with it. And it's not as privacy-invading as a video. So it's that combination of ubiquity, ease of access, ease of collection, the strength within the signal itself that meant we moved from being voice-video behavior to just being voice. The one problem we have had with voice and everybody has it in voice AI is it is so strong. It carries so much signal, not just about health, but about everything that or about everything around who you are as a human. So your age, your birth sex, your accent, your background, ethnicity, how many languages did you speak until the age of nine that affect how you sound now, where you live. So that is a very strong signal. And you need to differentiate that from what is the actual signal, say for depression, fatigue, stress, diabetes, differentiate out and isolate the signature that is a condition and that's kind of what we specialize in doing. Hermes Frangoudis (10:40) Very interesting and it comes multilingual. Emilia Molimpakis (10:44) Yes. So because we started multimodal and we have so many ways of covering the gaps, our models do work across different languages. We typically need to fine-tune potentially a little bit for a new language, but we've done it now across Greek, English, Spanish, Italian, Brazilian Portuguese, Indonesian and Japanese as well. So it's expanding now to all different languages. Hermes Frangoudis (11:10) And those are very, like when you think about the style of language of each of those, very Emilia Molimpakis (11:14) Yeah, they are. They are. Yeah. Hermes Frangoudis (11:17) intonation to tone to just style of speaking, right? Like some people speak really fast, just naturally. Some people are real slow, loud. Some people just always sound angry when that's just like culturally how everyone sounds. Emilia Molimpakis (11:27) Yes. Yeah. My co-founder Stefano always sounds angry. He's Italian and he's loud. He's angry. There's a lot of gesturing. And we can still tell what is happening. One key thing we do, which is also very important is yes, we kind of adapt across different cultures, but when we are working with users or individuals, initially we compare them to other people in the dataset who will match them or potential confounds like age, gender, and so on. Hermes Frangoudis (11:35) Okay. Okay. Emilia Molimpakis (11:58) But then over time, we actually build individual, longitudinal models per speaker. We do that automatically. But basically, you become your own baseline over time to how you speak. Like Hermes, how you are speaking. Like you have a deep voice. You're quite calm. But we'll be able to learn what is "normal" for you. And that's where kind of the real strength comes in. Hermes Frangoudis (12:21) It's really cool. So over time, my timeline becomes part of the baseline instead of just strangers. Emilia Molimpakis (12:27) Exactly. Yeah. So you are your own baseline over time. And that's very tricky to do. It takes a lot of longitudinal data. as far as I'm aware, we're the only company doing that at the moment. It's pretty, pretty tricky to implement. And we only require now, ⁓ just three samples. And I don't think I mentioned our sample size is very small. We only need 15 seconds of speech to do a full assessment at this point. It used to be 10 minutes. Now it's down to 15 seconds and it's Hermes Frangoudis (12:38) you. Emilia Molimpakis (12:54) real-time processing, so you get your scores immediately, basically. Hermes Frangoudis (12:57) Yeah. I tried one of the demos that we had moving around internally from, from Ben, an amazing demo and it blew my mind. I'm standing there at the train station talking to, going through the exercises and within, like you said, a few seconds, I had an answer as to like how I was feeling that day. And for the most part, lined up like, it was amazing how it caught that. ⁓ Emilia Molimpakis (13:21) That's great. Good to hear that. Hermes Frangoudis (13:22) So yeah, I would definitely recommend it. Like seeing is believing on this one. Emilia Molimpakis (13:26) Yeah. Hermes Frangoudis (13:27) So now that you have this dataset, you're on the journey, you and your co-founder kind of bet it all on building this set, this model. And like you said, it paid off. Can you tell us about some of these like real-world applications and how does it really deliver on these health and safety insights at scale? Emilia Molimpakis (13:46) Yeah, absolutely. So I think I mentioned like we kind of started with mental health expanded to other areas of health and healthcare was our initial starting point. So we work with large health organizations like the NHS, we work with pharma companies and there we are supporting them in assessing, monitoring, triaging for various conditions through voice essentially. So very kind of clinical application. However, in the last couple of years, I would say, one of our biggest areas has been outside of healthcare, we realized that being able, like having a model that can identify through a few seconds, whether someone is fatigued, or whether they're stressed or distressed, actually had very big applications in safety-critical environments. So environments where your fatigue or stress impacts your safety or the safety of those around you. So prime use case, one of our biggest areas is automotive. So We embed our models within vehicles to detect driver fatigue and stress in real time to help avoid accidents. You may not have it currently in your car, but the majority of cars or vehicles are moving to voice-first communication with the vehicle. And we kind of embed ourselves there. That's super big. And then other applications like ⁓ aerospace, airlines, construction, mining, defense, again, where safety is impacted by fatigue and stress. We basically help directly. More recently, it's actually become even bigger. So with the whole world moving towards voice-first communication in terms of not just human-to-human interaction, but also human-to-computer interaction, very soon you'll be talking to ChatGPT and Claude through your voice, not just typing. This is a huge opportunity for us to provide essentially a health and safety infrastructure layer for all of those voice communications. So we are enabling safe voice agent deployment at scale by being able to detect distress, stress, depression, panic, pain in voice to enable better interventions and catching people when there is essentially a need. So it started healthcare and now it's kind of like everywhere. It's very exciting. Hermes Frangoudis (15:54) That's amazing. the health and safety piece kind of is like, it feels like one of those like, man, like that feels so obvious, right? But it's just not. The fact that this technology exists and is making that like a reality, right? And because in a lot of these industries, are feeling like, I can't admit when I'm tired, right? Like, no one's going to believe me. But now it's like, hey, you Emilia Molimpakis (16:02) Yeah. Yeah. Hermes Frangoudis (16:19) probably should slow it down. This is not going to be one of those things you power through, right? Emilia Molimpakis (16:25) Exactly. And I think for us, it was very interesting and we needed to go through the healthcare space in order to get to the safety space because it created trust, like automotive manufacturers, airlines and so on. They trust us because we've worked with the NHS and healthcare. So it's created this trust kind of domino effect. Healthcare trusts us, therefore safety-critical trusts us. Safety-critical trusts us, which is also regulated. Therefore more healthcare people trust us as well. It creates this very nice trust loop. Yeah, it was definitely, it wasn't the obvious thing to do in the beginning, but now it seems so obvious. and it's, yeah, it's just, ⁓ I'm so excited that we can help in so many different areas. It's become something way more powerful than I ever could have imagined. And yeah, just very excited about it. Hermes Frangoudis (17:13) That's gotta be an amazing feeling as a founder. You know, the thing that you've thought for so long and had that like passion for is like, you're seeing really the fruits of it. Really in this time period where you're going through, how do I build up to this? And how do we get there? How did you really decide on like, would behavioral and speech signals were like most scientifically defensible, right? To gain that trust of the NHS and the healthcare industry to be able to like say, you know, it's valid, it's real. Emilia Molimpakis (17:40) Yeah. I think it's a combination of different things. Like one of the biggest assets we've had as a team has been the team basically. We, as a company, we don't have many salespeople. I'm the only person doing sales. We have a massive science team who's separate from AI/ML, separate from engineering and so on. And collectively the team has over 290 peer-reviewed publications in this space. So it's like really rigorous scientific background. And that's what enabled us to know what to look for to begin with. We were extremely cash efficient. We're a UK company. We're not based out of San Francisco. So our original rounds were not as large as some of the SF rounds. So being able to be efficient there was super key. And so, yeah, we based the original datasets and what we were looking for on that background, on that scientific research, and then the models were super strong. Then you go through, you know, certifications from cybersecurity data processing to actually regulation as a medical device. And all of that builds trust over time. And the more people that trust you, you know, the more that kind of creates that flywheel effect. So in the beginning, it was based on existing research. Now we've got to the point where our dataset is so big and so rich. We know everything about everyone we're gathering data on clinically, just making that very clear. Within the bounds of yes, everything is very, it's a whole, we go through ethics approval processes. We have an ethics committee review everything we do. It's very rigorous. Through that, we've actually been able to find signal in new areas we never anticipated. For instance, diabetes. If you told me six years ago, you could hear diabetes through somebody's voice, I'd be like, Hermes Frangoudis (19:02) within the bounds of. Emilia Molimpakis (19:22) "Are you crazy? That's not possible". It turns out it is. There's a lot like nerve damage through chronic hyperglycemia that is affecting the muscle coordination and nerves on your vocal cords, capillaries in your lungs, the fatigue in the lungs, meaning you can't breathe as much, like the capacity kind of shrinks. All of that kind of creates these signatures. And it's been like now going like strength-to-strength. Identify your own signal and then kind of building, validating, iterating, filling the gaps and kind of going from there. Hermes Frangoudis (19:54) It's like almost mind-blowing, like how the effects from all these different pieces all over your body, diabetes is something that affects it. Like a blood sugar issue, right? And so You never think that it also wears on your vocal cords. You hear of things where they're talking about like neuropathy and that nerve damage aspect, but the cascade throughout the body Emilia Molimpakis (19:57) Yeah. Hermes Frangoudis (20:17) which then manifests through the voice. Sorry, I'm just having this moment where my mind is kind of like Emilia Molimpakis (20:25) No, it's very powerful. In the beginning, I guess the key thing for us is making sure we do everything right and making sure we're being really rigorous. So initially it was like, are we sure this is diabetes? Is it something else? Are we picking up something else? Is it BMI that we're picking up instead, which is correlated? So then you have to actually gather more data to control for that. And it's like, no, it's not that. Okay, is it something else? Is it maybe the medication that somebody is on because they're diabetic? Is it something else? Is it something else? And so when you start to actually dig in, then you build more of that confidence like, no, it's actually, this is what we're reading. This is what we're genuinely measuring. And so yes, it can be used to support and to help clinicians and other people. We never intend to replace doctors with any of the things that we're doing. It's more to augment their capabilities and to allow them and others to focus on what they do best, which is kind of human-to-human interaction and so on. But yeah, it's extremely, extremely powerful. Hermes Frangoudis (21:28) That's amazing. You have like this very rich validation pipeline. We can call it, right? Like, cause you're validated. Okay. This signal is actually this it's not see. And so you're going from like, just collecting this data to being able to show the signals. And then is that then peer-reviewed and how does that kind of make its way through to again, cause you mentioned like, this is all through the research and through these cycles that build trust. Emilia Molimpakis (21:53) Yeah, absolutely. It's typically, it's very similar to how like an academic kind of process would look like just kind of lot faster and more efficient. start with a solid hypothesis, typically based on signal or literature, gather a large amount of data from the right population at scale with the right labels, multimodal, if you can control for confounds, iterate, assess in a larger scale, a kind of more representative real world sample. We've started complimenting the dataset with synthetic data to fill in the gaps. So we're actively now, creating data that like voice data, synthetic voice data that has a health signal in it. We're able to do that now. And that enables us again, to fill in more gaps. We build the models, test them out, productionize them, and then we publish on it. We're smaller team, so we can't publish as quickly as we'd like, but it does go through that whole process. It's peer-reviewed. Yeah, I mentioned earlier, kind of the whole team has loads of publications that they all want to publish. It's just, you know, it's a process. But yes, we're currently, we have papers being reviewed in multiple big journals, including like Nature. Scientific publication, so excited to see the new ones coming out. Hermes Frangoudis (23:08) That's huge. That's so exciting. Like not only are you pushing the voice AI space forward, you're pushing the scientific space forward in general and like multiple directions. Emilia Molimpakis (23:16) you Hermes Frangoudis (23:17) So you're in this academia mode, research mode, building the models. How do you bridge this gap now to like real world, right? Like NHS trusts you. Now, how do you bring this into the real world setting? Like what's the difference between like that data collection and actual real-world environment? Emilia Molimpakis (23:35) Yeah, I guess there is a huge difference between academic research and then kind of real-world research, productionization of models, application. The scale is completely different. So the dataset size is hugely different. And that brings with it a whole set of complications, like confound signal is amplified with larger scale. When you're in the real world, you do not have control over how people are interacting with you. Accessibility, everybody has kind of different devices you need to account for. They may have all kinds of comorbid issues, health issues and so on that typically in academic research, you control for variability in environments, human behavior in general, just not being predictable. Like that is one of the most amazing things I've found like when you do it at scale is people behave very oddly. You do not expect them to behave this way. Like one of my favorite examples of weird behavior is when we were doing voice and video in the beginning, we had a user who didn't want to show their face. But instead of not engaging or turning their camera off, which was an option, they decided to do the whole thing with a towel over their head. And they just interacted with us with a towel over their head, which muffled their voice. It created all kinds of issues. But there's loads of examples like that. So human behavior at scale is very odd. And then kind of things like privacy and security. Like you aren't controlling when people are gathering data or like, Hermes Frangoudis (24:47) Thank Emilia Molimpakis (25:02) speaking to you, there will be other people in the background that you may be picking up on. There's a loud environment, they're in the car, they're at a train station. Like you said, you were doing your demo at a train station. We need to account for all of that. All of that is like the variability has to come into the productionization. That's been one of the biggest learning curves, I think, for us. Hermes Frangoudis (25:23) I just have this vision of there's so many memes in the developer world of like how the developer envisioned it, how product envisioned it, how the user uses it. And I'm just picturing this person with like a towel over their face. Emilia Molimpakis (25:34) Yeah, yeah, yeah. It was like a full towel like over their head and they did it multiple times. It wasn't just one time. It was like a longitudinal data collection exercise and they just insisted on having a towel. I think we reached out to them. We're like, you can turn the camera off. It's okay. But yeah, that was very interesting. I think that's my favorite, but there's lots of other ones as well. Yeah. Exactly. Exactly. Or, we're here to capture a voice. Silence. I was just like, like in the beginning, it was prompting, like there were no voice agents back when we first started. So there wasn't a way of prompting people, when they were talking to you. Now we can say with a voice agent, like our own voice agents, Hey, are you there? Tell me a bit, blah, blah, blah. No, in the beginning, it was just like, do this. And it's just like silence. And then they submit the recording, you're like, okay, great. That's not gonna create anything. Yeah, human behaviors, very weird thing. Hermes Frangoudis (26:35) Very true. And that's one of the things you don't really consider when you first start any sort of venture like this. You're like, no, this is there's a need for this people can get it. It's gonna be so simple. But it's all right. You get over it and clearly that didn't stop you, right? So. Aside from crazy behaviors and not even crazy behaviors, odd, personal sensitivities, just quirks in general, right? Those aside, like from the early deployments of the models, things like the diabetes being like something that surprised you, was there anything else that kind of like surprised you on things that were being picked up or learned about? Emilia Molimpakis (27:15) I think, yeah, probably the new health conditions were the most interesting, unexpected things we were finding. Human behavior as well was unexpected. We found a lot of novel effects, I would say, in voice. Like, how do people compensate for loud environments? Like, when you're in a car, the way you speak is very different to how you speak outside of a car. And that means we need to adapt our models for that. So we deploy different models when we're working with cars versus the non-vehicle environments. Multiple voices and making sure we are able to lock-in to the right voice that has consented to us processing their data and giving them back health information or well-being information, I think learning how to deal with all of those has been the most interesting novel things for us. Hermes Frangoudis (28:06) It sounds very like, it's one of those things that you don't think about and then comes through and you're like, wow, this is really how we're gonna move forward and actually scale this. It's like you've cracked that nut now. Now you can kind of keep going, right? Emilia Molimpakis (28:15) Yes, it It took a while, but yeah, now we're kind of a lot more like a machine of, okay, this is how we gather data, this is how we process, this is how we build, this is how we It's great to finally be at that point. Hermes Frangoudis (28:26) So what were some early assumptions? Maybe that turned out different than you had originally hypothesized or thought that. Emilia Molimpakis (28:34) ⁓ Yeah, I think all of those different things that kind of I mentioned earlier. I think in the beginning we were also hypothesizing that we would need multiple types of data, like voice, video and behavior in deployment as well to capture the signal we needed. That ended up not being true. You can actually build super-powerful models and go with something that is very easy to get from people like voice versus say video and much shorter samples. As I said in the beginning, we needed 10 minutes of data, the very, very beginning. And now we've managed to shrink at the beginning of the year, I think it was 60 seconds. Now end of year, it's 15 seconds. I don't really think there'll be a lot more like kind of shrinking at that point. Maybe it would go down to 10 seconds. And then the kind of how quickly you process and feedback the data. Again, it used to be minutes. Now it's 0.2 seconds. You get kind of processed information. So I think the speed of change also is probably something we weren't necessarily expecting. Hermes Frangoudis (29:39) That's amazing. From like almost an hour to 10 minutes, right? Like that's a massive order of magnitude to already come down. And then 15 seconds, that's like. Emilia Molimpakis (29:50) It's super quick and it opened up all these applications. Like you can't use 10 minutes of data in a car that is then produced like. Yeah, you can't record 10 minutes of speaking data and then give them the scores back in five minutes. Like it just, that's not an application, but the moment it's 10-15 seconds in 0.2 seconds processed, then it's real time and you open up all these applications. Now we are Hermes Frangoudis (29:59) right? Emilia Molimpakis (30:16) at the point of providing outputs like biomarker enriched outputs on health, on safety and so on. And that in real time informs policy reasoners we're building to then recommend different actions that are adaptable in different environments. We work with the partners we're working with to create these policies. So that would not be possible if we couldn't do all this processing and recording in such a short time. So it's creating completely novel products we hadn't imagined in the beginning, like this health-and-safety infrastructure layer through voice in real time. That's completely different to mental health diagnoses, which is where we started. Hermes Frangoudis (30:57) And it's quite the path growth, right? Like the fact that you could go from such a long kind of recording cycle, which only really works in probably healthcare and medical situations, to now this bite-sized, almost like social media clip of voice can now kind of, Emilia Molimpakis (30:59) Yeah. Yeah. Hermes Frangoudis (31:17) it makes it applicable everywhere, especially now as voice becomes so much more ubiquitous as like the interface for communicating with the technology. Emilia Molimpakis (31:26) Yep, absolutely. Yeah, it's like it's less than a TikTok video. Not that I use TikTok much, but apparently it's a nice way of saying it. ⁓ Hermes Frangoudis (31:34) All the social media platforms now are really getting into that bite size format. So the fact that it fits within something that is guaranteed, right? Like the signal is guaranteed to have attention for 15 seconds. So you get a very probably good reading in that time period. What's the next 12 to 18 months look like right now? Like you have so many opportunities. What are the milestones that you're really looking forward to? Emilia Molimpakis (31:58) Yeah, I think this year has just been insane. It's just been incredible. I think not just for us, but anyone within voice AI, 2025 is the year of voice. And I think that is so true. And I think what's super exciting right now is this deployment of the different applications I was kind of discussing at scale. So we're very excited to see this health-and-safety infrastructure layer we are. Creating deployed at scale with large telecoms providers, LLM providers and more. So the deployment piece is super exciting, but also regulation. So I mentioned like, we try to do things like as well as humanly possible. And we don't take this lightly, like this is health-and-safety data. So we are actually the first speech biomarker company in the world that is getting regulatory approval as a medical device, which is a huge milestone for us. It's very difficult to do, but we're doing that and I'm super excited about and we're becoming Class II diagnostic device across multiple different geographies, which is very exciting and different conditions. The conditions themselves are again, exciting. As I mentioned, we are now diabetes is the one we're pushing for the most. So we're now actually running clinical trials and comparing against blood biomarker data and like kind of the real hard ground truth, the gold standard. And it's very exciting to see that's going to be coming out very soon. And then respiratory health data, cardiovascular health, not data, models and outputs. And also edge deployment is something very big for us. We're going from cloud, on-prem to on-edge, which completely changes what our partners can do. And I would say probably the last one would be funding. We have some very exciting funding announcements coming out quite soon. So yeah, can't say too much, but watch this space. Hermes Frangoudis (33:49) That's awesome. I want to get candid here because you mentioned some pretty interesting things and I want to ask you, so from a regulatory and compliance perspective, like for people that are interested in this sort of thing, what were some of the toughest regulatory or compliance questions you've had to address so far? Emilia Molimpakis (34:05) I would say one of the hardest things in general to overcome as a, I think any AI company working to become a medical device has to overcome this problem, is regulators in different regions don't all understand how AI works really in practice. And so for the longest time you had to deal with the fact that regulators were forcing you to regulate essentially a version of your model, which meant that, say I've got this dataset, I've built this model, V1, it takes these inputs, creates these outputs, and it's built on say 10,000 people, whatever it is. I regulate this as a diagnostic tool, has a stamp on it, and we move on. The problem with AI is that the very next day when I get even one person's new data going in, that version of the model is obsolete. You iterate, you build a better version because you should and that's how AI works. But regulators didn't see that. They force you to regulate a version, which actually is stifling innovation. It's stifling your ability to make better models and it pushes people away from regulation. So they would rather not regulate, but be forced to just maintain an outdated model and then pay again, do the whole process again for the next version. That was the biggest issue, I think, and challenge to overcome. We've managed to overcome it. And now we've got to the point where you can actually regulate not a version of the model, but what goes in and what comes out of the model. So provided you keep the inputs and the outputs the same, you can regulate that piece and you are able to iterate and update the actual model itself. I would say that was the hardest thing to get people to see, see it kind of come to fruition and actually develop and if that didn't happen, I think it would be extremely difficult to regulate our kind of technology. I don't know if answered your question. Hermes Frangoudis (36:00) No, I totally did. And I think that was very interesting. Like that, that initial hurdle of just, this is how the process works and this is how the AI process works, which was like. Emilia Molimpakis (36:10) Yeah, it's just like, yeah, I remember, like, I think I remember speaking at the EU Parliament at some point on this and saying, guys, I, like, I want to get stuff regulated. The way you're forcing us to regulate is not realistic. You need to adapt. And thankfully, things have changed. Like, there has been adaptations, which is really great. They're not amazing, but you know, it's not perfect as a process, but it's a lot better than it used to be. Hermes Frangoudis (36:39) the regulatory aspects and helping move that needle forward for everyone. That's huge. Emilia Molimpakis (36:43) Yeah, Yeah, we're actively part of multiple different cohorts right now, trying to improve how things are regulated. Like right now we're part of the Digital Medicine Society cohort, which is creating the new gold standards for digital phenotyping for mental health. Basically what are the digital signals you can and should look for mental health diagnoses. And that is hopefully going to affect how the new psychiatric manuals, DSM, currently it is the DSM-5, how the DSM-6 actually diagnoses mental health conditions and we're helping take that forward now. So it's a year-long cohort, but I love doing these kinds of things because, you know, this is how you move things forward. Yeah, that's really exciting as well. Hermes Frangoudis (37:30) You mentioned you got a funding round and I can imagine if regulatory and compliance is kind of misunderstanding some of this space. Like what are some of the things that maybe investors are not really understanding and things you'd have to educate on and would like to educate on? Emilia Molimpakis (37:48) Yeah, I think there's a, there's probably a lot. When we started with mental health, it was the whole mental health side that needed a lot of explaining. Now, I think with, with voice AI, people are a little bit more like investors are more up to speed with what this technology can do. The amount of investment that's gone into voice AI this year alone has shown that people are understanding it. I think I mentioned a little bit earlier as well, the key thing I see from investors is this underestimation of how difficult it truly is to build these kinds of models because of data and know-how, scientific know-how. It doesn't mean that because OpenAI or like these big companies have so much funding. It doesn't mean that they can automatically just build this overnight. Like we know we're talking to them. They can't, otherwise they wouldn't be working with us. So it's not about scraping the internet. That has lost value now. People have got to the edge of work, like to the end of what you can do with that. It's about human curated datasets, really rich labels, know-how and so on. And the amount of just effort that needs to go into that. I think that is something still investors are not all on board with, but yeah, hopefully they are getting there. They're understanding it slowly. Yeah. Hermes Frangoudis (39:07) Totally makes sense. I could see where there's people out there that think, well, know, Whisper AI is out there. Like there's all these other models. You're just wrapping and it's no, this started from scratch. This is very specially curated exactly for this purpose. Emilia Molimpakis (39:20) Yeah, exactly. Yeah, that's another thing is like every time I speak to an investor, it's like, yes, but what models do you use? And like, we built the models. No, no, no, no, no. But what do you use? ChatGPT? Do you use Claude? No, no, we built the models. Yes, but what like, okay. Let's start from Hermes Frangoudis (39:41) You gotta wonder how many times they've asked that question and the person's like, well, you know, under-the-hood it's actually. Emilia Molimpakis (39:46) It's Like, no, under-the-hood, there is no under-the-hood ChatGPT. It's us. ⁓ yeah. There's always a moment there. Yeah. Hermes Frangoudis (39:55) But that's got to be a great feeling. They're like, so what are you using? And you're like, I made it. Emilia Molimpakis (39:59) Yeah, we made it. Yeah, we, yeah. Yeah. Hermes Frangoudis (40:02) The team made it, right? Like this is ours. So aside from maybe misconceptions and investors and stuff, the voice AI is kind of like an exploding field in general and within health and safety. What do you think is maybe like an over-hyped trend and what do you think people are just like, aren't paying enough attention? Emilia Molimpakis (40:23) Yeah, I think with voice AI, like, you said, kind of like exploding, there is an over emphasis or overestimation of the value of automation, automating human tasks or automating processes that are already there. And don't get me wrong. I think AI-powered automation of things is fantastic. It should be done, but it's not the be-all and end-all versus say deeper understanding through AI or something deeper than a human could do. Like an example I can give you is kind of this intersection of voice AI with mental health, for instance, where you have loads of chatbots and voice bots now, where in mental health, they are simply automating questionnaires that already exist. But those questionnaires are subjective, they are biased, they are broken. You are automating a broken system. You're just making it automatic. You're not solving Hermes Frangoudis (41:16) So you're the breaks. Emilia Molimpakis (41:19) You are. And there is this false promise through say voice automation of mental health questionnaires. It's this false promise of increased access to care. You're not increasing access to care. You're increasing the number of people who then need to be seen without differentiating them and propagating an existing problem and creating more bottlenecks. Whereas if you actually say, focus on biomarkers or focus on a more deeper understanding and different way of looking at it, you can start to break that problem. For instance, that's one of the ways we work with the NHS is we don't just do like automating a questionnaire. We will figure out through the biomarkers at the point of entry who is truly suffering and who needs to really be seen very quickly versus who can wait a little bit longer versus actually this person isn't suffering from a mental health issue. It's something entirely different. Therefore, they need to be triaged out into a different pipeline or maybe they don't need to be seen by a doctor. They need to engage with, say, for instance, social prescribing, which is a thing in the UK where you engage with community. And a lot of the time people are lonely and they reach out to doctors. I think about a third of all doctor appointments are due to loneliness because they want to engage and talk to someone. Being able to differentiate and help there can actually create better flows and can address the problem rather than just augmenting an existing bottleneck, I would say. So, I'd say that's probably the most overhyped thing I'm seeing. Hermes Frangoudis (42:47) I mean, it makes sense. It's kind of the thing on everyone's tongue is like, oh, AI's going to take over this. It's going to automate everything. Emilia Molimpakis (42:55) Yeah. Yeah. I mean, AI is going to help a lot of stuff. It's going to automate a lot of stuff. But at the end of the day, humans are very good at doing certain things. AI can do other things. Everybody basically has their own expertise and specialty, and that will come out over time. Like, I'm not worried. I hear people getting really, really worried about what AI is going to do. I don't think it's not going to be that way. It's just going to eventually be filtered out in that way. Like, this is what AI is good at, this is what humans are good at, and it's complementary, essentially. Hermes Frangoudis (43:26) Makes sense and it's like you said, the AI is able to kind of go and think more deeply and process a lot more of this information in one go than the human could but the human can direct the AI. The AI isn't thinking of new novel approaches. Emilia Molimpakis (43:43) Absolutely. And I think it's so so important to have the concept of a human-in-the-loop. Like with our models as well, we have the concept of a clinician-in-the-loop. We don't want to replace the clinician. We want to augment their capabilities, but they need to oversee how it's used. An example, for instance, is we can detect say within depression and anxiety, we can actually detect and isolate individual symptoms of those conditions. There's about 15 different symptoms. So mood changes, fatigue issues, attention, but also appetite problems. So we can detect if somebody is having issues with their appetite, not eating, overeating. We can detect that that's there, but we don't know why. We don't know if this is due to depression or did the person maybe just have norovirus or did they eat something weird? We can't detect that difference, but a doctor can. And so the doctor needs to look at it and then ask the right questions that a human asks to be able to actually identify what is the real root. And human empathy is not something that a voice agent either can actually replicate that well. As I said, human interaction has so many strengths that AI wouldn't be able to replicate, they can just augment it. Hermes Frangoudis (44:55) I know we're getting to the top of the hour and I want to be very respectful of your time because you have been so generous with us in sharing so many insights. Emilia Molimpakis (45:04) Thank you. Hermes Frangoudis (45:05) And this one's a bit of a wild card. So if you weren't working with voice, let's say, what other part of the human cognition would you be using AI to understand? Emilia Molimpakis (45:15) That's a very good, it's a very good question. I think I'm very biased. Like I've always done language. So voice is. You can say language. Yeah, language is something I've always looked at. But I think because there's so many things that you can look at through voice, One thing I would like to look at would probably be just consciousness itself. Probably trying to understand, ⁓ Hermes Frangoudis (45:22) You can say language, you can say voice. Emilia Molimpakis (45:38) What is human consciousness? What is consciousness? Can we replicate it? Can we build it with AI? That would be super interesting. I think it opens up like a whole Pandora's box, but I think it's a very interesting area. I think now we're seeing like a lot of physical AI as well with robots and so on. That is going to be like the next big thing I think is consciousness. Maybe in a few years. Hermes Frangoudis (46:00) No, it's an amazing thing to just think about, just like sit there and pause. It brings a whole another set of ethics questions when the machine now has consciousness, but doesn't get a break. Emilia Molimpakis (46:07) Thank Definitely. Yeah. as I said Pandora's box. It's a whole different thing, but it's interesting. Intellectually, it's huge interesting problem. Hermes Frangoudis (46:19) Philosophically, intellectually, I think it's actually really cool and interesting space. Never heard someone talk about it and it's like, wow, that is, yeah, that'd be an amazing space. ⁓ Emilia Molimpakis (46:30) Yeah, I think it's coming. I think it's coming, yeah. Hermes Frangoudis (46:34) I want to say a very big thank you to you, Emilia and Thymia for joining us today. And to all our viewers watching live and to everyone listening along. Please like, subscribe, and we'll see you on the next one. Emilia Molimpakis (46:48) Thank you.