Welcome to Tech insiders where we speak with founders operators, investors and thought leaders in the education technology industry and report on cutting edge news in this fast evolving field from around the globe. From AI to xr to K 12 to l&d, you'll find everything you need here on edtech insiders. And if you liked the podcast, please give us a rating and a review so others can find it more easily. Martyn Farrows is the CEO of soapbox labs, the global leader in voice AI technology for
children. Martyn joined soapbox labs in 2017 as chief operations officer to lead operations, business development and strategic partnerships. He was appointed to the CEO role in May 2021. Founded in 2013 soapbox Labs is an equity centered privacy by design voice AI company that powers accurate and equitable voice experiences for children of every accent, dialect and stage of
development. By the end of June 2023, soapbox had powered the practice screening and assessment tools of over 50 global education companies and delivered more than 80 million learning experiences for young students soapbox is the first AI company to be certified for prioritizing racial equity in AI design by digital promise and the EdTech equity project.
Martyn holds a PhD in European politics, and has more than 25 years of experience in ad tech and AI for children with a particular focus on data privacy and ethical approaches to leveraging AI in the classroom. Dr. Martyn Farrows Welcome to Edtech Insiders.
Thank you, Alex. It's a pleasure to be here. And thank you for the invitation.
Oh, my It's great to have you here. Soapbox Labs is an incredibly interesting edtech product. Can you give our listeners a little bit of the story behind the founding of soapbox labs and how you came across speech recognition for children, you and your co founder as such a core goal for soapbox? Sure,
well, the founding story of the business is really around Patricia Scanlon, who is a kind of veteran, although she might not like be calling a veteran speech engineer, I mean, worked in the industry for 20 odd years, right? And started her own family and was actually observing her own daughter using educational kind of literacy apps on an iPad. And she came to the realization that these apps were not actually capable of measuring the expressive voice
of her daughter, right? So she was playing these games, and she was effectively gaming the game. So you know, in terms of measurement of progress, it was very hard to really take anything meaningful out of that. And being a speech recognition engineer, she said, Well, okay, well hang on a second, why is it that nobody is really tackling the problem of children's speech and being able to assess that? So she said about trying to
solve that problem. And that was the kind of genesis for soapbox labs,
voice recognition technology has been on the rise for years, but that children's use cases really been under the radar. And it's obviously so important for education, use cases. Can you give us just a couple of ways in which voice is currently used in ed tech products that is now powered by soapbox?
Yeah, I mean, it's a really interesting sort of study, as opposed in terms of the history of voice recognition. Because when you think about kids, I mean, there's always an assumption that you could just take an adult set of models, and make it work for children somehow, right? There's two big problems with that. One is that the data set that you're using to build the adult models doesn't include children's voice data. Now, you could say pitch change that voice data and make it sound
like children. Right. So that's one thing you could do. However, you know, kids language behaviors are very different. Very different, you can't mimic that side of it. Right. And the second point is, which is really the question is the use cases for kids are so different, right, particularly in education, right? So you're interested in not just what did the child say? But how did they say it? Right? And also, in many cases you're interested in, what
didn't they say? Right? So you're not just sort of like if you take a you know, the most obvious use case of a voice assistant in the consumer world as a as a voice recognition in the consumer world is a voice assistant. voice assistant is trying to kind of like, define intent from what you're saying, and then give you a meaningful answer. But that's not really that relevant in education, particularly when you think about literacy applications. So the use cases are very diverse
and varied. Then the other thing is, you know, kids, in terms of how they use language, they're not a homogenous group. Right. So a two year old is very different to a four year old is very different to eight year old. And then the use cases are very different across those
different demographics. So Now we've the product working in many, many different use cases, you know, from the whole sort of literacy journey, so it kind of preliterate kids can have fun, logical awareness, all of that kind of stuff right up to reading fluency and comprehension. But even things like screeners for dyslexia, and so on use cases, very diverse and varied and very diverse and varied within the demographic that we deal with.
Voice technology is such an amazing way to scale certain aspects of especially literacy instruction, that we just don't have enough experts in the world to evaluate every kid all the time, the way
we would want to. One of the things I find so interesting about soapbox labs, just for listeners who may not be familiar with it is that it's a, it's really a B to B product, it is the underlying technology that does all of the really high level speech recognition that underlies many publishers or other companies, speech education tools, can you tell us a little bit about that model,
it's kind of a deliberate strategy on our part, which was to say, like, we are very, very good at what we do, right, which is to process voice data from kids, we're very, very good at that. But we're not in any. And we wouldn't consider ourselves to any degree education experts, or assessment experts. But there are lots of really talented and creative folks out there and businesses out there that have been doing this for decades, and who are very, very good at that
stuff. So our kind of vision and our mission was always well, let's work with the best people out there who really understand teaching and learning and assessment. And let's work closely with them like hand in hand with them and a partnership model, where we can support what they're doing. And we become the sort of, to borrow kind of marketing phrase, the Intel insights are where they're sort of the enabling technology. But we don't want to make any judgments about what the
technology could be useful. And you know, even when we started on this journey, we had some ideas about how it could be used. But as we've gone through the journey, our clients and our partners and teachers that we work with come up with amazing applications for the technology and ideas for how you could use
it differently, right. So it's a symbiotic relationship, which is a win for both ourselves, our clients, but also the Teaching Learning Community and the kids themselves, because they get the benefit of that sort of confluence of expertise.
Absolutely. And that type of model, I think, is not as common as it is in other industries in edtech. But it's incredibly powerful, because it allows everybody to specialize. One company that I always admire is learnosity. They focus on assessment, they have like 70 different types of assessments, but they are embedded in other publishers and all sorts of other products. And I think there's a similar model there. Let's talk about the
technology itself. One of the focuses of soapbox Labs is it you know, obviously, accuracy is incredibly important, but also equity, and privacy. So tell us about the equity challenges in speech recognition, technology, and sort of how you think about the whole sphere of what you're going to do with this voice data and how you make sense of it.
Yeah, and it sort of comes back to your earlier question around your use cases. And what's unique about children. I mean, one of the unique things around working with this technology, particularly in education is that it's high stakes, right? So you know, you can't afford to get things wrong, right. And you can't afford to get things wrong for any child that's potentially using the technology. So it's going to work equally well for all children. But that's not really the case with a consumer
product, right? Because if you, you know, end up ordering the wrong pizza with a voice assistant, it's annoying, but it's not high stakes, right. But if you end up incorrectly grading a student by giving incorrect data, because you haven't understood that their accent is different. That's a problem. But so we very early on thought about this, you know, in the privacy world, there is a concept of kind of privacy by
design, right? And we took that concept and thought about in terms of equity by design, right? So in other words, as you are designing the product, not the user use of the product necessarily, right. But as you are designing the product, what are the actions that you are taking in order to ensure that it is going to be equitable once it becomes productized? out the
other end? Right. So fundamentally, that comes back to this idea of how representative is the data that you're using to train the models of the cohort of children that are going to use this right. So you've got to make sure you have representative data. But you've also got to ensure that that representative data has been collected from kids who are using it in an authentic way.
Right. So it comes down to really granular things like so if you're asking a Spanish speaking child to describe a picture, show them a picture, which is culturally appropriate to them, right? Because then they will use their authentic voice to describe it because it's something that they're
comfortable with. Right. So it's that level of detail around you know, ensuring that in the design process You are taking steps and put in place processes and systems that ensure that you are aiming to mitigate any bias out of the system.
So that I mean, that's incredible. And obviously, you know, so box has been working for many years to support, accents, dialects, diverse linguistic backgrounds. You know, as you say, the soapbox labs, it works within dyslexia screeners. It works for aligning with dibbles, and meta metrics, Lexile and you know, grading, you know, high stakes assessments for oral reading fluency, all of these really important assessments in the
classroom. How do you know, if the soapbox labs technology is fairly grading and not accidentally taking students with a particular accent or dialect and systematically, you know, grading them in a different way?
Well, I guess our approach really, though, is to be agnostic and objective, right? So our role is not to say, what is the correct pronunciation of a word, right? Our job is to correctly build a system that is able to recognize the multiple correct pronunciations of a word, right? And then it's, then it's for the educator on the other side, or the curriculum developer or the assessment provider, to take those objective data points and use them to measure progress
that that child is making. So we're sort of providing a reliable and valid set of data points, which are in and of themselves, objective, and then when applied to a rubric, have meaning in terms of how that child is progressing, or what measurement is being used in order to assess their progress.
Gotcha. It reminds me a little we've talked to some plagiarism detection companies, and I think they talk about it in a similar way, which is that they don't want to be the final arbiter of, you know, cheating or not cheating, what they want to be is a very high level detection system for behaviors, which can end you know, for, in your case for linguists for phonemes, so that you can pass through incredibly accurate data to the end users who are actually going to be deciding on the assessment
exactly
results. Yeah, yeah. How do you deal with noise, you know, in educational settings, classrooms, that can be very, very noisy, that you could probably hear other kids in the background sitting next to this kid, how do you tackle the issue of noise interference when it comes to accurate speech recognition?
Yeah. And that's, it's a really good question. Because when Trish started out in this journey, you know, there were there were some speech recognition systems out, which, you know, did work to some extent with kids, but they always required the QSB. Wearing headset mics, right under the audio was really clean, because the audio had to be cleaned for the system to be able to understand what they were
saying. But anybody who's a parent or a teacher will know that kids don't generally operate in quiet environments, right? So right, if you're going to build a system that works in the real world, your data set, and this comes back to the underlying data set has to be representative of the real world that those kids want, you have to build that into the models.
And you have to train the system to be able to ignore cross talk or extraneous noise in the background, a TV going off, or if you're in a cafe or in a classroom. And there is there's obviously there's a threshold, beyond which a system will not be able to distinguish, right in the same way that the human ear, you know, if you're in a really noisy environment, sometimes it's really hard to figure out what people are saying, there's
always a threshold. But as our system is very effective at working with those environments, because the data that we've used to train the models, has been representative of those environments.
So Fox has many years of experience recording data in classrooms, through all of these different partner organizations that do various types of speech recognition. And it feels like what I'm hearing from you is that corpus of data, real world data from classrooms, you know, kids from different linguistic backgrounds, all sorts of things, is really a major differentiator for you compared to anybody who are to say, try to start this kind of company this year.
Yeah, it is. And I think, you know, you mentioned that diamond is coming up on 10 years now that we've been doing this, right, so we've probably got coming on to maybe 100 million sort of interactions with our system at this stage, so many 1000s 10s of 1000s of hours of data. Right. And that's really, really important. And that going back to your earlier point about as being a kind of b2b infrastructure provider, you know, everybody benefits from that, when you think about it,
right? Because of the scale, the kind of data network effect. Everybody that uses the system benefits from the fact that it is the most representative and therefore most accurate and most equitable solution that's out there. So that is a key USP of ours, if you want to call it that, but it's also you know, there's an awful lot of knowledge that's been gained over the years as well in terms
of how best to do that. That's right, and how best to do that for kids, as opposed to how to do that for adults, or how to do that for the market. And that also is something that, you know, is a huge benefit of the business, you know, even to the point of when you're collecting that data. And when you're using that data for your training models, how do you annotate it,
right? Your annotation methodologies are very different for kids speech data in an educational context, and that you might be for, you know, adult speech data in a voice
assistant context. So there's a lot of sort of expertise and know how that's been built up over those years in terms of how to use the data or things like, you know, simple stuff, like, how do you elicit a four year old to use their voice in a educational app where you're asking them to kind of pronounce a word, you know, what, what are the mechanisms that you use to trigger them to start to speak and to stop speaking, you know,
those kinds of things? Because it's very different doing it with a four year old than an eight year old, right? So I like those interesting learnings that we've, we've developed over the years,
I can imagine I'm sure people are nodding along with the idea of how do you get a four year old to stop speaking? That's the universal question. You we've mentioned a couple of the education partners that you have, but we've been speaking in a very abstract way. Maybe we could talk through just to you know, have our listeners really understand exactly how this is working and what scale it's working at. What is the range of partnerships? So box Labs has formed over these 10
years? Are there specific examples or success stories that are either wide ranging or unusual that you'd love to share about how soapbox has really made an impact?
Yeah, I mean, I think Trisha is sort of vision at the very outset was this notion that in order for this technology to have impact, you had to get it into the hands. To use a metaphor, there's probably not quite appropriate, but into the hands of as many people as possible, right. So to get as many of them interacting with it as possible.
You know, to do that, you have to be able to operate in a market where you have everything from, you know, companies like scholastic and McGraw Hill that have been around for decades and serve 10s of millions of kids right down to, you know, much smaller providers who are probably digital first entrants into the market and a building with voice in mind from the outset. Right. So we have everything within that sort of
spectrum. And I think what's exciting for us is that, you know, when you see the end user product, right, and you see the impact that that is having. So you know, with McGraw Hill as a product thing, I'm reading mastery, which is for kids who have reading challenges, and it's an intervention product, right, so you're seeing how this technology can be used to support encouraging those kids through their literacy journey with the new product from Scholastic, which is ready for
reading. It's really around those foundational literacy skills. And it's combined with the amazing IP that scholastic have around printed books as well. So it's not a purely digital product. It's a blended product, which is designed to fire the imagination. And then you look at, you know, companies like Lynn Gumi, which is a UK based company, but operates largely in East Asia, a lot of kids in Taiwan, two to three year old kids in Taiwan, using Ling Gumi to learn English.
Right. But we're so Buck's power in that experience, so very, very different sort of use cases and product design decisions that have been taken, but all of them, you know, making amazing and engaging use of the technology.
Yeah, it's exciting to hear when you mentioned, you know, get it into the, I don't even know get into the ears or, you know, get the correct phrase. The whole time, I was like, What is it, I don't think there is a really good way to say that, but it's fantastic that you know, Scholastic is I think over 100 years old now.
And it's so amazing in this moment in, you know, tech history that we have the level of AI I mean, when the first Siri came out series named after the Stanford Research Institute, Sri, that developed it, and you know, that felt like kind of magical at the time, the idea that we could have accurate voice recognition, and at the time, it wasn't even that
accurate. Now, we are at a very high level of accuracy, being able to combine that level of advanced tech with things like a scholastic books or McGraw Hill readers, or Global English learning, it's sort of an old meets new paradigm. How do you see that, you know, panning out in the world? It's such an interesting way for things to come together.
Yeah, I think, you know, for me, the exciting thing is the complementarity of it, right? So it's kind of like, it's where you see that intersection of the technology and, you know, more traditional media, but also, where people are looking at the technology and thinking, Hmm, you know, how could we do things differently now that we know that we have this technology that works and is accurate and is dependable and reliable? And what does that mean? In relation
to things like instruction? What does it mean in terms of practice, because practice, particularly around literacy skills is such an important thing. And what does it mean in terms of assessment, right? How does that change the paradigm and how we think about those things? particularly where we have a situation which I think everybody in every, you know, economy around the world recognizes that, you know, teachers really valuable resource, but also quite scarce,
right? Time starved, right. So their most effective interaction with kids is when they're having that kind of one to one interaction, but that's time consuming. You want to take away things, for example, that might be taking up time and allow them to sort of spend more time on the things that are most effective and impactful. So the technology can help in those
things. So you know, rather than having teachers administer an assessment or a test, you know, maybe the assessment can be done in the background, by the technology. And then the teacher has just surfaced with the relevant information which they can use to design interventions, which would give them actionable insights and make best use of
their time, right. So there's, there's lots of exciting kind of uses for the technology that are starting to emerge that go beyond sort of where I mean, most technologies when they come into the market, they used to automate the way that we currently do things, right, because that's the obvious thing
to do. And so how do we automate how we currently do things, but then you move quite quickly to a position where you're saying, actually, how do we do things differently now that we have to we have this technology, where I think he's most exciting? Have you
seen any of that sort of self assessment type of behavior with some of your partners, where instead of students feeling like they're doing a high stakes exam or doing, you know, even a reading fluency assessment, it could feel like you say, describing pictures or having a conversation or, you know, speaking in a very casual way, which still generates useful data?
Yeah, you're starting to see a lot more of that. Now, I think, you know, what excites me the most is sort of closing the gap between instruction practice and assessment, right? And they're becoming a sequential become a linear process, the model is a
circular process, right? So we are starting to see, for example, you know, in a reading practice context, right, the way kids are practicing reading at home, right, so they could be in the home, that the teacher is then being able to see the practice that's taking place and see the effectiveness of that practice. Right. So that's a really valuable insight that the teacher will get that they wouldn't necessarily have if the technology weren't there, in the
first place. Right. So you are starting to see those applications. But it's always, you know, the education industry, for good reason, I think, is generally quite thoughtful about how it applies technology, right? Because it's high stakes, right? So we have
to be thoughtful. And we have to be careful about how we introduce this technology, we have to build confidence in usage of it, we have to be very clear about the purpose for which we're using the data and, you know, hold data privacy piece, we have to be transparent about that, in order to be able to achieve the benefits and the impact in the long run.
That is a perfect segue because I wanted to talk about the privacy piece of this puzzle. We mentioned it sort of in passing earlier in relationship to equity. But you know, we're in a very strange moment with AI tools everywhere and people sort of a new cycle of people giving data to services without entirely knowing what's going to happen with that data. That is incredibly important for school.
So tell us about how soapbox Labs has handled the privacy of voice data collected the ownership of that data, compliance with school regulations, how do you think about privacy as a company that has, you know, your ears on so many different kids?
So very early on, we engage with an organization called Primo, which is an FTC approved himselves a safe harbor, whatever the phrases that they use, but they basically have a program for COPPA and GDPR compliance. So we get audited every year to ensure that we are compliant with any privacy regulations, not just covering GDPR, but they're the primary ones. And obviously, as COPPA and GDPR, you know, as they evolve, you know, they also evolve their, their auditing
processes. So, so under, sort of regulatory side, we've ensured from the beginning that we are compliant, but we've also taken it a step further in terms of kind of like, well, well, actually, we want to be thought leaders in this space, and we want to effectively establish a benchmark for how this should be done, right? Because, no, don't forget, when we started this 10 years ago, you know, certainly over here in Europe, I mean, Alexa didn't exist, right? Then
Alexa came along. And that became the sort of the reference point for everybody around how voice was used, and in a lot of ways, very positive. But it also in some ways negative around the data, privacy aspects of it, right? The weaker kind of saying, well, actually, no, we do it very differently, right. So first of all, we rely on consent as the legal basis for processing the data, right? So we don't process any data unless we have the consent to do so the
necessary consents to do so. The second thing is that we are, you know, very, very transparent about the purpose for processing that data, but we're very clear that we're only using it for the purposes of delivering a speech recognition service for kids that is, in this, let's say, an educational context for helping to improve literacy, right? We're not using the data for, for any marketing purposes, we're not going to microprofile the kids, we're not selling that
data to anybody else. We are specifically and only using it to improve the accuracy of that system. But then we go a step further and say, well, actually, we never want to know the identity of any of the kids that are sending voice data, right. So everything that sent to us is anonymized. And a lot of cases
also de identified, right. So we're actually sort of, it's an entirely different model around privacy, because our sort of raise on debt for now, our mission is around building an accurate speech recognition system. It's not about collecting information from consumers in order to be able to target them with ads around products that they might like, based on the preferences that I've inferred from the things that they've said, right? So
it's right. It's those principles, which having a really important, so you know, consent, transparency, and then there's this notion of, you know, anonymization, and never actually wanting to identify an individual child.
That's a very thoughtful approach. And I appreciate you calling out that sort of the commercial use cases, the Alexa is of the world, the series, the, you know, voice assistant use case is so common, and it's so ubiquitous, that it's become sort of the paradigm in which people think about this, for better or worse, right? And
wrong with that, as everybody is. And if we all buy into that, like, I use Siri, you know, I mean, the series probably not good example now, because Apple do take a slightly different approach. But I use Google, and I use my voice to ask Google maps where I want to go and I allow them to collect that data. I use Alexa, sometimes, you know, it's kind of like an AI. But I'm going into that knowing what to do
with that information. But what we're sort of saying is, well, actually, you can flip that entire model on its head, that doesn't have to be the standard, right? So with kids in particular, you have to turn and say, Well, no, this is this an entirely different way of using
that voice data. And we shouldn't be, you know, necessarily applying the same sort of principles to kids, as we do to adults, we have to be a lot stricter around how we build these kind of systems, but it has to be conscious and deliberate. Alex, right. So you know, it's not like privacy by accident, or equity by chance, right? Be deliberate and
conscious about it. So we building these systems you put in place are the systems and processes that make sure that privacy and equity are at the heart of it, and everything then sort of flows up from the bottom as opposed to hoping that it all works out. Okay.
It's a great point that, you know, children don't always have the ability to consent to data collection. And if soapbox is used in apps in homes, then it becomes you know, some of the same issues that you know, an Alexa has in a home and you know, all sorts of things can be picked up on a microphone, it is very important to have all of these different
aspects in place. And I'm sure that your, your partners, you're probably very used to explaining this, because I'm sure your partners are also very interested in how to make sure that they are not on the wrong side of the regulations, or some horrible news story about somebody picking up the wrong data.
It's an interesting point, if I could just interject, because, you know, in the majority of cases now, particularly where we're working in the US with clients who are delivering into schools and school districts, we actually don't retain any of the voice data, the data sent to our system, we process it, we send the result back, and then it's deleted, because the school districts require that of their clients, you know, so it's actually there's a lot of misconceptions around, you know,
there's, again, going back to the consumer world, we're told, in a lot of cases, that we need to keep your data because in order to do that, we'll give you a better service. Actually, that's not true. Because once you've reached a certain level of accuracy, and so on, you don't necessarily need to keep all of the data, they want to keep the data for other reasons. In our case, we don't need to keep that data, reprocess it, we send the results back, and then we delete the voice data immediately.
Yeah, I think it speaks to the value of thinking through the power of speech recognition and voice recognition, in a totally different context than the consumer one. You've mentioned this a few times, you know, when you're speaking to a voice assistant, you want it to do something for you. So you know, close enough could be close enough, you're it's an adult voice speaking, you expect that, you know, the Amazons and Googles of the world are going to do everything they can to
collect your data. We've all had that experience where the phone serves up an ad and you're like, Hmm, you know, I was just talking about that. I wonder if it was listening to me. And it creates this sort of whole different relationship to it versus an educational use case. You're looking for accuracy, equity, privacy. Absolutely don't Yeah. So it's really interesting to hear you outline the differences.
So the common thread across both of those is the utility of the technology, right. So we know it has utility, right? Because it delivers value, we can see that but in order for us to take advantage Attack utility in education, we have to think differently about how we build the system that delivers that utility. And that's really
important, right? So we don't want to pass by the opportunity of that utility, because we haven't understood that we can actually build things differently in order to address a lot of the concerns that we might have around consumer
products. So if we take another example, like early bird education is one of our clients who have built a dyslexia screener, that dyslexia screener has the capacity to be delivered at scale at low cost for children to be diagnosed with potential reading challenges at a much earlier age than hitherto we were able to do. Right. So the benefit of that in an educational context, not just for the educational ecosystem,
but for the kids themselves. And we play a small part in that way, you know, the early bird product is amazing. And soapbox engine is a small part of that. But we are enabling that to happen, right? And we don't want to lose the utility of that, the opportunity of that, by not understanding that we can actually build these systems in a safe way that actually does respect fundamental rights of kids.
Absolutely. I'm curious, you know, we're in this incredibly fast moving age of generative AI and you know, new products and people introducing new features to their existing products is sort of fast as possible to incorporate this new technology. How do you see voice enabled tooling playing a role in this wild, you know, AI fueled moment? Obviously, it is artificial intelligence that makes voice recognition happen?
How does it sort of connect with some of the things that you're seeing in AI within these existing ad tech or new ad tech companies?
Well, the first thing I would say is that, you know, voice recognition and voice technology as an interface technology is valuable, right? So it doesn't matter what you're interfacing with. So but the existence of genuine of AI now is another dimension that we can interface with, right? So there is no we're already seeing, for example, the ability to be able to respond using voice for that
voice data to be processed. And then that plugged into a generative AI LLM that generates the next question that the child wants to respond to based on the initial voice response that they gave. But so the feedback loop suddenly becomes much quicker. And when you think about things like personalization, and so on, that's a truly personalized feedback loop, because the next item that they're being asked has been generated, based specifically on the response that they gave to the last one,
right. So that's, that's a really interesting application and voice as part of that feedback loop, and becomes a really exciting part of that feedback loop, because it's the interface technology as opposed to a keyboard or a touchscreen, or whatever it might be.
And it's an interface technology that is much more accessible to younger students than typing and reading, you know, so, you know, we talked about conversational AI, and what you're describing is true conversational AI, it's actually speaking, and getting spoken back to and, you know, I have not seen that in action.
I'd love to see that. It seems like a, you know, as you describe it, it doesn't seem like the hardest technical problem to do a voice interface connected to agenda AI, you know, LM on the back end, but I actually haven't seen it,
use it, we'll see it I think over the coming weeks or months, I know that a couple of pints of ours are working on sort of prototypes of that kind of technology, right. So, but from our perspective, the important thing is that the voice interface piece is accurately reflecting what the child said, right. So that it still works, right, that that part of it works on the LLM side, on the generative AI side, it has to be an appropriate knowledge domain
that has been used. Right. I know you did a an interview with Santia, from Merlin mind. I mean, they're doing some really interesting work around what is, but it's actually an appropriate knowledge domain to use in education that is not hallucinating, that is giving back content that is appropriate for the context that you're using it in. And I think that's the exciting thing about generative, generative AI and lmm is, what it's delivered to us is the art of the possible,
right. So we now know that with these massive knowledge domains and information that we can extract meaningful information and generate new content on the back of that, right. But now what we need to do, particularly in education is to take a slice of that and say, Well, what's what's relevant and appropriate to a five year old who's doing history in a classroom in Utah right now? Because they don't need to be accessing chat GPT to be getting those answers, right. So I think those things are
really interesting. But again, the voice interface part of it doesn't go away. And it has to be accurate. It has to work for all kids, and it has to be, you know, consistent and reliable.
It feels like that model of how do you take the concerns which are legitimate and address them in a way that's true, you know, really actually address them. You've been doing that for voice recognition for 10 years trying to make sure that people don't that say, Oh, I bet you're gonna start selling things to my kids. You're like, No, we can prove Have through 100 ways that that is not going to happen. That needs to happen for generative AI very soon.
Yeah. But that goes back to my point around, you know, looking at the design process, as opposed to, we tend to focus when we're thinking about concerns with these technologies on the use of the product. Right. So if we go back to the equity conversation, it is about the AI design process, and how do you build equity into the design process?
That's where for me, the regulatory environment should be looking at the design side of it and saying, Well, what data have you used to build this model, but show me where you have, you can demonstrate that what the data that you have used is representative of the target audience that you're going to be
working with. Right? Don't don't just show me that you've taken a pre existing data set that you don't really know if it has any inherent biases in it or not, but you've built their system, and you're hoping that it's going to work for these kids, but show me that you've actually taken time to curate a data set that is representative, because then I can be confident that the end product is actually going to work those kids, right, and that's for me where it needs, we need to focus on the design
process itself. And not just the end product
100%, we've seen just in the last few weeks, the ASCD, which is, you know, giant, nonprofit educational training group in the US that just acquired the technology group that just merged the technology group put out what
they're calling a chatbot. For schools, it's called stretch, and they're doing, they're trying to do exactly what you're describing there, you know, get a model where it's totally transparent, where everything's referenceable, where the data going in is, you know, reliable, all the things that you would need to safely use something in schools, they're trying to, you know, jumpstart an initiative like that. And I anticipate a few different people are going to really be jumping into that,
including oil in mind. And then combining that with voice. Yeah, removes another kind of bias, which is the bias towards text communication, I mean, young children or people who are functionally illiterate, which is a huge number of people can't use Gen AI as it is. And it's really interesting to think about the possibilities there.
Yeah. But again, you know, it comes back to this principle of, well, let's be thoughtful about how we look at this, right. And let's actually drill into one of the systems and processes that have been put in place in the design stage, as opposed to just looking at the and that's a really interesting example, I hadn't seen that. And I think, you know, you will see more
examples of this. Now, you know, it's the concept that we can actually take these massive datasets and use the sort of LLM approach to build meaningful interfaces for people is, is, in principle, really exciting. But now we have to think about what's the application of that, in a context like education, which, as we've already discussed, is high stakes. And we can't afford to get it wrong.
100%, I loved your phrase earlier, the art of the possible, you know, it opens up all these possibilities. But it also opens up all these fears, people and I personally, anybody who listens to this podcast will know I'm on the optimistic side. I'm very
bullish on this. But I think the possibilities have to outpace the fears, you have to have some incredibly positive outcomes that are really exciting, sort of mind blowing and memorable before something happens, it's going to be mind blowing and memorable in a bad way, which is, you know, going to happen as well. So it's an interesting moment, I want to circle back to one more point that you made earlier that I thought was
really interesting. And I think it's relevant to this part of the discussion, which is, you mentioned closing the gap between instruction and assessment. And I think this is something that's been really a dream of educators and instructional designers for a long time, because assessment raises, I mean, assessment can feel very formal, it can get
people very anxious. And I'm curious, you know, how you might put together this sort of technological stack we're talking about in a model that's actually formative, you know, that where a kid is talking to a virtual tutor, or friend or you know, pet, and the things they're talking about, become the assessment data, and you can serve up, hey, we noticed that the child is having trouble with a sound or we noticed that they're mixing up these two vowels. Do you see a future like that?
I do. And I think, you know, I've always been fascinated by this notion of measurement, right? And how do we measure progress with kids? Right? And we've typically always done it by designing tests, because that was another obvious way to do it. And like, you know, I'm not a psychometrician. I'm not an
expert in assessment. But you know, what, I've spoken to a lot of people who are, and their dream is to be able to say, well, actually, if we could just have assessment in the flow in the background, that is not a test. Nobody likes it and tests, but it was the practical way to
measure. Right. But there are other ways that we can measure I think, and I think that that's, again, you're starting to see that happen that you have that kind of assessment happening in the background in the flow, particularly for young kids, right, you know, that like getting five and six year old kids to sit, literacy and numeracy assessments is now Not necessarily either the best way to measure them or the most productive and effective way to think about, well, what are the
interventions that are required in order to be able to help them move along in terms of their progress, right? Because it's too lumpy. Or you just, it's just too, you know, I can do at the beginning of the end of the year, you've missed, you know, so many opportunities to intervene in the intervening period. That makes any sense. So it's kind of like, now we're
starting to see it. But I think one of the things that we are working very hard on is understanding the data points that we get back, and what does that mean in terms of designing the future frameworks of measurement that you might use, right, but part of the challenge there is that you need longitudinal datasets to help practice that process, right?
Because, and we've always had a very strong connection within the academic research community to work with partners, like the Florida Center for reading research, for example, because those guys are amazing in terms of thinking about how do we do things differently? And what does that mean in terms of policy decisions that are taken in the future? That's not our job, but what we can do is to help them with the data that would support that research?
Absolutely. And the underlying infrastructural technology that could enable all this dream of continuous embedded voice enabled formative assessment, it's this, you know, it's never been possible before. And I feel like we're getting really close to it. It's an exciting moment.
Yeah. So as we come on the end of our time, I am very, very curious to hear your answer to this from your particular perch at soapbox labs, looking at AI looking at voice, what is the most exciting trend that you see in the EdTech landscape that you think our listeners should keep an eye on?
For me, exciting means the impact, right. And I think the most impactful thing is a lot of the work that's going on in the assessment space now. And look, it's probably not the most exciting, you know, like for a lot of people because, you know, measurement and assessment. And tests have traditionally not
been all that exciting. But actually, when you think about the impact that that could have on individual kids, if we can design more effective assessment that is, you know, in the flowers, you certainly talk about, you know, formative, it's
the area. And I do think generally I will have a huge, I do know of work going on in the language learning space, for example, where they're using generative AI models, in order to be able to design what I was talking about earlier, where you've got input using voice, the system, the assessment is happening around that input, that's the person that the learner has given, and then they're generating the next item just on the fly using a
generative model. I mean, that's the true sort of example of a personalized learning experience within language learning. So, you know, for me, that's where the excitement is, is in those kinds of spaces. I do think that, you know, we're doing a
lot of work now. And maybe if I saw the announcement this week, you know, we're gonna build a Spanish language version of our engine that's exciting for us, because, like Spanish speaking kids, a huge cohort in the US, but also globally, you know, to be able to build multilingual versions of the product is really exciting.
I hadn't heard that announcement. And that's a huge deal. Because it is a very widely spoken language by many, many students in lots of different countries, including the US, that is a very big deal on,
it was huge, kind of thanks to the support we're getting from the Gates Foundation for that, because they see those opportunities and their cohorts of kids that they're working with. And it's, it's a real, really good example of how you can partner in that sort of public private way to drive innovation in the marketplace that we're so keen on.
It's really exciting. What is a resource that could be a book blog, Twitter feed newsletter paper, that you would recommend for anybody who wants to dive into the topics we discussed today? A little deeper.
So I mean, the obvious, I would say is like the Starbucks labs blog, because there's some really exciting stuff there. But I think the law material that's produced by the getting smart guys are particularly around the the practical use of AI in this kind of like demystifying what it means in education. They're really good, digital promise resources are amazing. So if I was to go anywhere, I think that getting smart guys and digital promise outside of soapbox, so the places I will go to.
Fantastic yeah, as always, we will put links to those resources in the show notes for the episode. Great suggestions. Dr. Martyn Farrows soapbox labs, you enabling voice recognition in ed tech, all throughout the world. Thank you so much for being here with us on Ed Tech insiders.
Alex has been a pleasure to meet you and talk to you and hopefully we may be again to get to meet in person at some point in the future.
I'd like that a lot. We'll make our way to your neck of the woods soon enough. Anytime. That'd be wonderful. Thanks for being here with us Dr. Martyn Farrows soapbox labs. Thanks for listening to this episode of Ed Tech insiders. If you liked the podcast, remember to rate it and share it with others in the Ed Tech immunity for those who want even more Ed Tech Insider subscribe to the free ed tech insiders newsletter on substack