Hello and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. When the idea of doing a podcast episode with SoundHound came up, the name rang a bell for me as an early music identification service.
Turns out SoundHound Inc. has become much more than that.
Writing about its Series D funding round closed earlier this year, TechCrunch said SoundHound is leveraging its 10-plus years of experience and data to create a voice recognition tool that companies can bake into any platform. and that they're poised to become a third neutral option to Alexa and Google Assistant.
Some big names and some big talk. But then again, the company's now valued at a billion dollars.
So I guess we're talking about a unicorn.
Sanhan Inc. has partnered with some of the biggest names in the auto industry, as well as Motorola Mobile and Sharp's robotics division.
Full disclosure, NVIDIA is also a strategic partner and investor in SoundHound as well.
Here to tell us how SoundHound has grown from a party trick app to a major player in the voice-driven AI business is VP of Product Marketing, Mike Zagarsic.
Mike, thank you so much for joining the NVIDIA AI podcast.
Thanks, Noah. Thanks for having me. Why don't I turn it over to you?
And I realized I said party trick, and that might have disparaging connotations to some people.
I did not mean it that way. You know, one of the earliest examples that I can recall of an app that recognized music and other sounds.
It was super cool, but it's grown into this huge player in the AI industry.
So How did that happen? Can you tell us SoundHound's origin story?
Yeah, it really is an interesting story because in many ways we're still a traditional startup.
But we also have a 13 year history, which in Silicon Valley is somewhat ancient. in some respects.
And the way it began, our co-founders were all Stanford graduates, two of whom basically wrote a very well-regarded PhD and actually in voice recognition and speech recognition.
And that was really always their vision.
And they wanted to commercialize that, build a company around it.
In 2005, Starting the company, they got a lot of interest, as you can imagine.
However, when they said, well, it's early days.
It'll take about 10 years to develop. Most of the investors said, great.
But is there anything you can do in the shorter term?
Because they believe in long-term vision, but obviously cycles are much shorter term.
So the idea of, of sound and audio recognition being derivative in some ways was born.
And what that allowed the company to do was build a business, scale, monetize while the underlying voice technology was being developed.
And then in 2015, uh, our new platform in addition to, uh, our soundhound product was houndify.com. the voice AI platform, and then Hound, our voice assistant, built off of it.
So we're young in many ways because... We've only really introduced the voice technology, but we're also old in some respects because we've been around for a decade plus.
Not to take us off track, but that sounds like an unusually healthy business growth strategy.
Yeah, I appreciate that. I've been with the company shy of three years.
So I had a good outside perspective coming in and having been around a few companies, large and small, When I heard that, it was eye-opening.
And what it's helped us be is somewhat more mature and really leverage... a long-term view.
So the recent rise to prominence is really a function of been driven by the funding and the strategic partnerships that have come about.
So we're we're in many ways, you know, a three year old startup, but also a 13-year-old at the same time.
And that's actually really helped our growth strategy.
Right. How big is the company? How many people are working?
Under 300 and continuing to grow. So if you ask again in a year, I'm sure you'll get a different number, but it's...
Right. Let's talk about the years since, well, the years since you've been there, basically since, uh, 2015. you said was kind of when the new era began and you've got the original SoundHound product the Houndify platform, and then the Hound Voice Intelligence app.
So what can you tell us about how those work, what they are?
Sure. Our founder and CEO, Kayvon in particular, really foresaw this evolution.
He knew it was coming and he had identified that if you want to create a voice interaction platform, you have to build it from scratch.
It's less effective to take something like tech and trying to adapt to voice, if you really want to create something unique, you build it from scratch.
So one of the unique properties of our Houndify platform is the data that powers it is very much voice data as opposed to text data.
And what that allows is, you know, we have a branded term for it, speech to meaning.
What it's doing is as the user is speaking.
It's processing what the user is saying in real time to extrapolate what's being said.
So it allows for corrections during speech as opposed to having to interpret what the user is said after the fact.
And that comes from an engine that is tapping into data, again, that was designed for voice.
So that's a big differentiator from the platform.
What are the advantages that that gives you, whether on the server side, the business side, or the user side?
Well, the simple one is just speed and accuracy.
Speed because there's no interpretation step afterwards.
And accuracy because... It's been self-correcting throughout the dialogue.
And then Hound is our voice assistant, and that is available on iOS and Android in the U.S., and it's a very helpful product for us in addition to being a consumer app that has a core loyal following.
It gives anybody a good glimpse into how Houndify works.
It helps us build our models and strengthen them.
And I think another unique facet of our organization is we're not providing just the tech or the platform.
We're building on top of it ourselves. The benefit there is, and this is, I think, a theme that I've seen through voice AI in general, is that you can't just produce the technology and let other people figure it out.
If you can't build an experience on it, then Why would you assume other people can't?
So Hound is very powerful for that as well.
It's sort of like a next level view of dogfooding in a way.
Exactly. Yeah. I'm looking at the Houndify page, as we talked about or I talked about in the intro.
You're working with a bunch of automotive partners.
There's also Motorola, and I come from a background covering the mobile industry, so I immediately started saying, hey, Moto, in my head.
And you're also working with Sharps Robotics Division and some other companies there.
Let's talk a little bit about automotive because that's obviously an area.
A lot of action in the AI space and people think about self-driving cars and that kind of thing.
But voice, a big deal in automotive. My family has two cars.
One is an older one with an old stereo, but I still connect my phone over Bluetooth and I say, hey Siri, all the time.
The other one's newer and it's got a head unit with, you know, some mobile phone integrations.
But same thing, using voice more and more in our cars.
How does Houndify platform work? SoundHound and AI and deep learning in particular, you know, what's the impact on our driving experience now and kind of looking forward?
Yeah, that's a great question. And what's great about automotive is that bring up that theme of short-term value versus long-term roadmap.
Exactly. So the industries that are going to want it now are the ones that are going to get immediate value.
And so obviously the hands-free eyes on the road experience is, I mean, as excited as we are about self-driving, we're far away from that.
So distraction-free driving is important.
Mm-hmm. And cars are becoming smarter and more connected.
Most car companies realize that it's a computer on wheels.
And if they don't adopt, I would say a connected, smart strategy, then there will be obsolescence as a result.
The idea of being a connected car is very much in everyone's roadmap.
And once you have the processing power, the connectivity, and the ongoing desire for safe driving voice is a natural uh, platform for that.
Right. And the, the, the short term value is things like car control.
So it's roll down the windows or, you know, how much oil, when's my next oil change.
But then there's the broader use case, which taps into our houndify domains, which could be stocks, weather, navigation directions.
So a strong cloud connection is really important.
And then underlying that, there's the broader extensibility of our platform.
So the short-term value is The driving experience augmented with data, but then obviously limitless potential beyond that.
And is it a similar approach you're taking with partners in other industries?
I would say generally is. One thing that we tell folks and our founder and CEO, Kayvon, is really clear about this.
He says that every company should have a voice AI strategy.
And whether or not you act on it is one thing, but In many ways, you don't have to go farther back than 10 years to talk about mobile and having a mobile strategy.
And the companies that had a mobile strategy were quicker out the gate and ready to act versus the ones that didn't have a strategy, had to figure out what to do and then act on it.
And our innovation cycles are are condensing.
So it feels like it's every 10 years is a wave.
And then likely every five, We don't have to really pull out the history book to see what happens to companies that aren't ready for the next evolution.
So that's something that we promote. And therefore automotive is moving quicker because they tend to have longer roadmaps.
They are really used to working with partners and it's a strong use case.
So it's really for us, the announcements have been a function of which company is ready to come forward and when.
Let's shift, it's not really shifting gears, it's related, but kind of go to the technical side of this a little bit and talk about deep learning.
How does deep learning power? Why is it important to voice and sound recognition?
And how does it play into the opportunities that you and your teammates see for SoundHod going forward?
Sure. I mean, the one thing that's become clear is that it's affecting every industry. even beyond having a voice strategy, if you're not thinking about how AI and deep learning affects your business, then you're already going to be behind as well.
And every organization has their own view on it.
For us, It's always that combination of algorithmic and human and manual.
So right around when it comes to speech recognition. getting data, processing data, our ability to leverage models to make that more efficient, rely less on human interpretation is key.
But I would say it's an evolving thing. I'm sure you've talked to many guests about this.
It's very much a philosophical approach that's backed up with real algorithmic solutions.
And then it's really a question of how and where and when do you apply them.
So I think Speech is absolutely privy to that.
But at the same time, because language is so fundamentally human, striking the right balance and And not relying exclusively on one, say, technique versus another is going to be important how we evolve.
How many languages, I'm putting you on the spot, but how many languages does SoundHound support right now?
So we're basically we're investing in a dozen languages.
So we obviously our our launch was within English, but we've now.
All of these partners that we've signed on, particularly automotive or global players, including Motorola.
So who are in. you know, say the 20 major markets with 12 languages.
So we are essentially actively developing all of those and have, um, basically made progress across the board.
And it just becomes a question of, you know, depth and sophistication.
But those are our main focus. Are any languages easier or more difficult to work with when it comes to natural language understanding and your speech-to-meaning technology?
I think it depends where you start, because we're the the key thing is that there's there's two traditional components to voice interaction.
There's the ASR, which is the speech recognition, and then the NLU, which is the natural language, as you well know.
And so. Language models in and of themselves are complex and every company is still fine-tuning them.
So if you take, I say, a tonal language like Mandarin, where you have four tones, there's a level of subtlety that makes it more challenging than single-tone languages.
So I would say any company that migrates towards the East and some of those more tonal languages will find those to be a little more adventurous, but I, in many ways it's, it works the other way around as well.
So that's, that's probably a, a standard reality for anybody.
You said tonal in reference to languages, but it made my brain do a little bit of a non sequitur jump.
What about in recognizing music and non-language audio?
Anything jump out at you from your time with the company as, you know, things that were particularly sticky to get the systems to learn to recognize.
It's really about the data that we get. So if we have good data and it's of a level of quality, then the bar algorithms do a really, really good job of matching it.
In the early days, it was really operational of getting the music data, getting it uploaded, making it accessible, making it fast and finding ways to optimize.
I think those are really the tricky things.
But if you have good digital data and you have a good underlying platform, then that part is straightforward.
Everything else is really about making It's the classic case of taking a great technology and productizing it.
Under perfect circumstances, you can do it, but no user is imperfect.
They have background noise. They may be talking.
They're in a loud environment or the sound is distant.
So all of the ways that our human ear can very easily filter that out has to be done at a technical level.
And those are often the sticky points that turn a product into into something people want to use because all of those edge cases have been addressed in a way that make it reliable.
We are talking today with Mike Zagorczyk.
Mike is the VP of Product Marketing for SoundHound.
They've been around, as he was talking about at the beginning of the episode, been around for some time now and had this really, what I think is a really great business model that hopefully there's more of these cases out there that I'm aware of. where the long-term vision kind of has always been to do what they're doing now with voice recognition turning speech into meaning and using AI to drive, you know, voice as the interface for the computers of tomorrow.
Those are my words, not Mike's. But in the meantime, they put out this really cool app that could recognize songs by hearing them.
And that kind of made them a name that stuck around certainly in my head.
And so it's very cool to hear about the evolution in the past three years.
In particular, as they've started to become a big player in the AI space with voice assistants and automotive and more, But Mike, let's shift gears for a second, no auto pun intended, and talk about your background.
You've been involved in Interactive for a while now.
How did you get started and what can you tell us about your own journey that led you to where you are today?
Yeah, it was definitely a process of, He did his commencement speech at Stanford and he had this three stories.
And one of them is. You don't always know how things are going to look.
You can only connect the dots in reverse.
And I've done that where... I started not knowing in college whether I wanted to go into something like computer science or marketing, which are two very orthogonal fields.
But I've always been fascinated by people, how they think and what they do.
But as are we. world is becoming increasingly technical, you couldn't separate what people did with how technology was evolving.
And I started in advertising. I thought ads were cool.
But I landed, of course, in the digital department where banners were a thing in the late 90s.
And so I was always the digital advertising guy before anybody really wanted to do that.
Are you claiming responsibility for pop-up ads?
I absolutely not. No, it's my job was really at the early days of advertising was there would be the Internet Advertising Bureau is just developing standards like. you know, standardizing banners and I was, I was staying on top of it to put that out.
But then I started working on, so I was at an agency in Boston and our client was Volkswagen and famous for the Drivers Wanted campaign.
So I was privileged to work with some really creative folks.
And that was really extending itself onto the web.
So it was these flash microsites, which everybody um fondly remembers or not depending on what it was but we did no pop-ups And my experience working in digital and building interactive experience brought me to Apple in 2007. where I led Apple.com and email marketing globally for the company from 07 to 12.
And Apple, of course, as everyone knows, has really mastered that balance of technology and interaction, right?
Steve would often say it's the intersection of technology and liberal arts.
And so I think that spoke to me, but I was very much on the receiving end of it.
And what I decided to do after five years at Apple.
Because when you're working at Apple, you're very much, you're isolated in many ways, just based on the company you work for.
And those were 07 to 12 were pretty, pretty busy days.
I mean, not that Apple doesn't have busy days, but those were, that was the iPhone.
That's right. My first day was when the iPhone launched.
Oh, my gosh. Most people – and at the time, it was a Mac and iPod company.
Yes. you know, put out this sort of interesting device that nobody really knew what was going to happen with.
And so it was. And of course, 2011 is when Siri launched.
I was there. Right. And it was interesting to be deciphering the value of a very early technology and what is obviously now referenced as a pioneer, but with a lot of mixed feelings.
Well-spoken, well-diplomatic, yes. Yes, exactly.
Early first mover advantage with all the thorns that come with it, I guess.
So that's when I broke into smaller organizations and worked.
Really, I developed that human-computer interaction mindset where... my people orientation, because I'm not a technologist, but inherently fascinated with how people act, think and interact.
And, and I realized to summarize where things are now, it's, It's the challenge we had an opportunity to have now is to take technology and productize it, right?
The best technology opens possibilities, great products solve problems.
And what we need to do is really balance what's capable with what's needed.
And I've found that there's there's a lot of room for that.
And I enjoy being at that particular intersection.
That's a better segue than I could have come up with.
So I'm going to run with it. What's needed going forward?
Uh, and obviously I'm not going to ask you to speak to anything proprietary that Sunhound's doing, but you know, to what you can't speak to or just broader strokes, you know automotive we talked about, and that's a great, great place to look, but in, in anything related to voice and computing and, you know, human-computer interaction.
What do you think is needed? What are the problems, opportunities that could lead to great products in, you know, the next, couple of five years, 10 years even?
I think it's a combination, we'll call it direct and parallel.
So within voice, one of the other reasons we're seeing traction is the three things that we offer, which is your own wake phrase, you know, your brand and your data.
So, What I think people are starting to figure out is when I say, hey, Mercedes...
I'm talking to Mercedes. I'm not talking to Google or Alexa.
And the wake phrase is a very powerful human reference point that sets the experience beyond that.
The idea that you get a wake phrase and the expectation is set there is very powerful, and that's really needed.
Not to pick on your former employer, but...
No, it's just to underscore your point. When my kids don't like the song I have on in the car, they just start yelling, hey, Siri.
And that's kind of, you know, they know, right, that that's how they, it's that human element for sure.
Yeah, it becomes a bit of a curse word. Right.
Brands are really picking up on the fact that as powerful as some of these consumer-facing assistants are, they don't They don't want that to become the dominant interaction for those brands.
So the wake phrase is really important. And so what needs to happen is, constant evolution of that core voice experience and essentially just improving the domains of knowledge, the breadth of what it can do.
But if it can do the core use case well, it'll find a market.
So in the car, if you have... great car control um you ask question i mean you know voice enabling the car manual i think is It's something that everybody's waiting to do.
It's like the classic, what does this light mean?
I don't have to flip through 300 pages. And that's useful.
And then in parallel, I think the thought-provoking stuff that I certainly like to drift on is making it more multimodal.
When do you get a voice response? When is it a screen response?
And what are some of the other cues? Is it gaze detection to say, well, only pick up when you're looking, for example?
Or how do we continue to add, I would say, complementary technologies that start to reflect the ways in which people want to react to make it seem almost that much more invisible.
So that's a whole chapter that I'm looking forward to seeing happen.
I think that's a great point, because I know I'm guilty of this.
And I think sometimes we all, when we're talking about, voice and voice is such a big thing in AI.
So we talk about it here a lot. We forget that a lot of times if you ask Your phone, for instance, you speak to it with your voice, you're going to get a response on the screen in certain cases or ultimately to get the information you're looking for. you know, it is a multimodal experience.
No, that's a fascinating thing for me to think about.
So I can only imagine in your day-to-day life, the challenges you're tackling.
That's right. If folks want to learn more about what SoundHound is doing, some of the cool stuff that you guys offer with your apps, but also what your partners are doing, Where should they go look?
True. I mean, our website is soundhound.com and that is the hub of everything we have.
And as we start rolling out more and more products are becoming in market, that'll be a place where we start developing more content and showcasing what's actually happening.
The other is equally important, if not more, is houndify.com.
Because that is our developer product. So users there can go and create an account if you have a A developer bent a free account.
It takes just a couple seconds to sign up.
And there you can really see what's enabled through the product at this point.
As well, so I know you likely have a lot of technical focus listeners and creating an account just saying on top of what features we're releasing.
That's, that's the great hub to do that.
What kinds of things are open to devs right now if they want to start leveraging your platform? build the client, as we refer to it, and start enabling domains and providing some initial phrasing.
So we have over 100 domains available with more coming.
And they can essentially turn those on and start building clients and products essentially that leverage our platform.
We are investing a lot more in allowing developers to do more custom commands and building custom domains.
That will really be the next chapter. But ultimately seeing... how it works, building core voice capabilities, leveraging the domains and understanding, say, what can a weather domain do versus stocks and whatnot.
It's a really good both learning development and product experience.
I mean, they can develop on it. And the developers that are with us now are eagerly looking forward to our roadmap as we continue to add more and more so-called self-serve capability.
Right. And collective AI and your collective AI alliance, what is that all about?
Sure. What is, I think, really powerful about where we are today and where we're going is Our founders, in particular, Kayvon, our CEO, he understood that speech recognition was only as good as what information was available.
And he always felt that the power of collective human knowledge is the best representation of where we are today as a planet and society.
So the underlying platform of collective AI allows for sharing an extensibility of domains, which is unique to us.
And that means that if I build a domain and I want to make the information in my domain available to someone else.
I can do that without any development required.
So therefore, if I put landmarks on a map, right, we'll say something like the space needle in Seattle.
The space needle is an object that has coordinates and it has a map.
I would say, so we'll call it a map corollary.
Therefore, I can say something like, show me the best burger joint within three miles of the Space Needle.
Right. And the reason I can do that is I didn't have to code anything about where the space needle was.
I just had access to that information. And so as people start developing more on the platform, once we truly open that up, the ability for it to understand reference points and leveraging information will create exponential growth across domains, both the linear growth, which we're seeing with siloed skills and actions.
And are you finding that your partners are pretty open to that idea of sharing?
Definitely. We will build custom and proprietary domains.
So if you're talking to Mercedes or... Honda or Hyundai, that's something that they'll make available for themselves.
But as if they choose to make information. information available more broadly within the Houndify platform, they absolutely could.
The benefit is they control it. If you put a bunch of information, you build domain, and you want to turn it off, you can.
So it's really an underlying knowledge problem. graph based platform right that is controlled by the people who contribute to it as opposed to one single entity.
And that roadmap is something pretty compelling and people really want to be a part of that.
Yeah, no, it sounds very cool. Is the answer, by the way, to that hypothetical question, is that Dick's Drive-In?
That's really the only burger joint in Seattle I know.
That's right. We'll have to pull up Hound.
All right, perfect. We'll do that after.
Mike Zagorsik from SoundHound, soundhound.com and houndify.com if you want to check out the dev tools and the platform specifically.
Thank you so much for your time. I appreciate the conversation.
And AI is at the tip of everyone's tongue.
And it's very... At this point, it's fairly segmented.
And I think for me and us as a company, the human side of it is always really important that you can't just build technology and expect it to work.
It's understanding how people think, act, and behave.
And how you build an interface for that is going to be key.
So what we're seeing is that we're not at a point where you can isolate a solution to one particular technological application.
And what voice I think is really unlocking is this need for I'd say deeper connection in terms of how we act with our technology.
So the idea that we can help foster that and bring that together and partner with companies and developers to do that, I think is just a really exciting time.
Well said. Best of luck in all this really, really interesting and ultimately quite practical work that you're doing.
Thanks very much. It's really exciting. Thank you.
Thank you.
