 So we're here to take current disrupts in London for you. My name is Injali, I'm an Enterprise Fellow of Royal Academy and also a Research Fellow of the University of Southampton. So here with the Royal Academy of Engineering, what are they helping you in the Enterprise Hub? Yes, the Enterprise Hub is basically a charity to help young triffiners like me to build up my business also for my PhD research. So where are you doing your PhD? Well, I'm just interested in this area of multimedia artificial intelligence. It's in... In... In Southampton University? Yeah, in Southampton University. Southampton University? Yeah. And who are you? Hi, I'm Mike Wald, I'm Professor Mike Wald. I'm Young's year's PhD supervisor and he got his PhD a few years ago. And he worked on Synope with me and the whole idea of Synope was a way to make teaching and learning more accessible. So we took lectures and recorded lectures, we used speech recognition to transcribe the lectures. So what is special about your way of doing speech recognition? So we can train our speech recognition model based on the domain knowledge and we can gradually improve the speech recognition result. So this is AI? Yes, this is artificial intelligence, yeah. So this is not just like... What is the standard way of doing this? It's comparing stuff, sentences and stuff, but you are doing it differently? Yeah, so we... The current minimal viable product that we are working with is universities for university lectures. So after they take the lecture, we'll give it to us and we will generate the transcript for them. So you take the lectures and then you can generate the transcript, how good is the transcript? It depends on lots of things. My lectures tend to be more than 90% accurate. So I use a good microphone and I tend not to speak too fast. Obviously it can reduce from that, but we have a method to correct the errors in the speech recognition by collaborative editing by the students, so we can get the accuracy back up to 100%. Nice, they can just log in and change a few errors that are wrong? Exactly, but the trick is they're all logged in, we're recording every single edit everybody makes and we have algorithms to compare the edits. We can look at the reputation of the student for how well they correct and we can give scores and rewards to students based on their editing. So it's super important to have transcripts of all the lectures, right? Well, there are lots of reasons to get transcripts. One is, in theory, it's a legal requirement to make videos accessible, particularly for students because they're paying customers, but also there's disability legislation. But also you can search the lectures from the transcript, you can translate them in different languages and you can actually read much faster than you can listen, so it makes everything a lot more useful. So you search for a word and then you get exactly directly to that place in the video? Exactly, exactly, so you can find things, but you can also bookmark and tag the lecture as well. Do you have a platform here in mind? Yes. So what is this platform? So this is our user interface, a collaborative editing. This is one of my videos, right? Yes, exactly, so that's one of your videos. So we can play the video and you can see that the transcript will be here and if you spot any problem in the transcript, you can sort of edit it. Nice. I said you can put this in the washing machine. Oh, this one is pretty accurate actually. So we can go for another one probably. You just need to get connected online. And we can see all the transcript down here. Yeah. We also have a print friendly option. So if you're replaying your video on a mobile device, there is a problem with the bandwidth you're paying for the bandwidth. So you can print it out and we have a QR code which means you can use your mobile phone and replay the video from that point. You can scribble and take notes on it and it makes the whole thing flexible. So how many are using it already? Is it real out there? We've been using it for about eight years at our university and there's some other universities around the world but up to now it hasn't been available. Well, it's been available for eight years but it's been more of a research project so we've launched this company of Sinnoh to be able to make this available to a wider number of students. This could be for anything, no? Everybody needs transcription of everything, right? Yeah, it could be anything. It doesn't have to be lecture videos. Our specialization currently is trained for the lectures but it could be used for conferences, for coaching, for video, for the real-time chatting. Or it could be used when people walk around conferences taking videos and interviewing people. I totally need this. Yeah, you do. So YouTube has some transcripts but how would you say maybe you compare with them? Are you better? Yeah, we are better. Roughly we are about five to ten percent better than their results. Which is very important because every time you hear or you see a mistake you get kind of discouraged about the thing and you kind of, maybe you want to, you know, forget about the idea and stuff. Yeah, as you just said, we can take the slides and information about the topic and train the system which YouTube doesn't do. How do you train? How does that work? Because in my video, for example, I interviewed a Swiss guy and you clicked on Swiss German or something and then it improved. Yeah, so there are a couple of problems about the accuracy of the Swiss recognition. So one important one is you don't have the right vocabulary to actually to describe what you're trying to say. So, for example, if you interview a company and if I get all the information about this company online, then probably some of the words will appear in your interview. So in this way, the speech recognition result can be improved. So, for example, the word synote doesn't exist in most dictionaries and therefore the speech recognition has to look it up in a dictionary. If you find it, it'll find the nearest word it can find. Whereas if you train it with the word synote, it will then be able to find it. And you should be able to recognize accents, right? Yeah. And then get into the accent and get better recognition automatically. Do you have to select it? Or can that be improved? Yeah. What can you do in the future? Yeah, definitely. I mean, we want to have a fully trained system that is easy to be trained by different people with accents or even their own vocabulary. So that's totally possible in the future. But also, when you correct the transcript, you can use the corrected transcript to retrain it, and so it can get better. So what are you doing here at the TechCrunches for? What are you looking for now? You start up, right? What's the next step? What's going to happen? So we are looking for different customers and ideas that the system can be useful. And we are also looking for financial investors who will be interested in investing in our technology. So this can be done here in the UK? It's fully done in the UK. Where are the people doing this? Well, I'm quite confident that we have fully done this in the UK. I've got support from the Royal Academy of Engineering. So hopefully that should be far away from us. So the best speech recognition in the world, potentially right here? Yeah, right here. Right. Cool.