Ricardo Herreros Symons, Chief Strategy and Revenue Officer at Speechmatics | BotCast #13
Meet Ricardo Herreros Symons, Chief Revenue and Strategy Officer at Speechmatics, who joined the BotCast #13 to answer our usual questions on building with Voice AI.
Who are you?
My name is Ricardo Herreros Symons. I'm the Chief Revenue and Strategy Officer at Speechmatics, one of the founding team. And what am I building? Well, we've been building speech-to-text for years, for voice agents. That's anything on the understanding side, whether it's diarisation or getting those words out and transcribed as effectively as possible, across multiple languages and accents. For us, accuracy is the most important thing.
How did you end up using voice AI?
Being in a speech-to-text company, it was always evident that once voice AI arrived, I'd be working in it. I think the moment for us was when our machine learning teams sat everyone down in the kitchen and did a talk on PaLM, one of the first large language models. As soon as that happened, we knew that speech was going to be the interface all human beings would use going forward. So that was the moment. We built voice agents ourselves. I'll talk later about how that evolved, but it was just the natural progression from where we were at that point in time. And now, of course, it's everywhere.
A moment of joy and pain using voice tech.
The moment of pain is interesting. We've always focused on having the absolute best accuracy and latency: the most accurate words at the lowest latency, and we knew we were really good there. But the moment of pain was realizing that there was more to it for voice agents. We looked at every single Pareto frontier and we were bottom left. It was great. The problem was, we hadn't focused on forced end of utterance, voice activity detection, end-of-turn latency. It meant that we went out there with a product that worked for some use cases, but for the first six months, a year, we weren't really landing it, because we'd been focused on different types of industries.
The other side of that is bringing all of that capability into a really simple and easy format. When Deepgram brought out Flux at VapiCon last year, they'd made it all simple and in the same place, and we probably lost six months catching up. The good news is that we've now got all of those things. It's really accurate. It's working great. But that was painful, not quite predicting how that was going to happen.
I think there are probably two main moments of joy for me in building with voice technology. One of them is, I've just got married, and I have an eternal goal to get our technology into my wife's life in every single way possible. She turns on the TV: it's our captions there. She uses a meeting note taker: it's us doing the note taking. Voice dictation: it's us powering the dictation. She reads earnings calls online: again, same sort of thing. So being able to touch everybody's lives and make them better is one joy.
The other one for me is the Arthur C. Clarke quote: any sufficiently advanced technology appears to be magic, or something like that. When I was demoing the first voice agent stuff amongst friends and family, I'd bring out a voice agent and they literally thought it was magic, because it was so advanced. Obviously, in the industry, we know what that's like nowadays, but being able to genuinely show them something that's never been seen before was pretty cool.
What lesson would you share with other builders?
I think probably a couple of lessons here. One, benchmarks. They are fantastic. I'm so glad that so many companies are doing them. It shortens the effort we need to put into testing models. But ultimately, always, always, always test with your own data. Make sure that it's going to work for your use case because, again, one of my favorite quotes: real stupidity beats artificial intelligence every time. You have no idea how a human being is going to break something. So until you've tried it out, tested it and been sure of it a million times, it's still not quite going to work.
The advantage of that is it means there are continual problems to solve, and that long tail, for me, is never going to be completely solved. But bear those two things in mind and I think you'll go far.
Where do you see the industry in 12 months and in 5 years?
In 12 months, I think the recent releases around speech-to-speech have been quite interesting, the recent Google and OpenAI releases. I imagine we'll see more of that getting towards regulated industries. It's not there yet, and on the hardest regulated benchmarks I still think cascade is what we'll be seeing. But that conversation will come up more. I also think we're going to see a lot more multimodal. You'll have seen the recent Tavus video where they've got a multimodal Turing test. Really impressive. I can see that really taking off. And then multilingual too, so multi-everything, to be honest: more things going on at once. People are going to demand that all voice agents are multilingual, because people need things to be simple.
My take on five years: I think it's pointless to be making predictions for five years, but I'll give it a stab. I think we're going to get into this really interesting world where, for the next year or two, we're going to be building loads to optimize agents that can understand anybody. However, in five years' time, I think the agents are just going to be so integrated that they'll be talking to each other. So actually the voice side of agents won't matter. The agents will be talking to the agents, and what will really matter is how we can create a very, very natural, very instinctive, very personal relationship between us as human beings and our personal agent itself. So I think agent to agent will not use speech, and the question will be how we create the best agents for personal situations. That's probably the most likely thing that I think happens.
On the other side of that, I also think the physical AI side is going to take off massively. I know Elon's talked about a billion robots. Unfortunately, his predictions usually come true in some capacity. So getting stuff on device, on really small form factors, private, is going to be really important as well.
SPEECHMATICS
BotCast powered by https://bespoken.ai 🫶
