Episode 25: Davit Baghdasaryan on the Science of Sound in a Distributed Work World

Howdy everybody! Today we're going to talk to the co-founder and CEO of a company whose technology makes it easier for those of us working from home to hear each other, even with all of life's noisy distractions going on in the background behind us. At Automatic, we say communication is oxygen. We're advocates of anything that makes communication easier and more effective, and one of the tools I find myself recommending over and over again is Krisp, which is an app that uses machine learning to mute background noise in just about any communication apps you use. For Krisp, David Bagdarossarian, there's even more to the story. He's leading a young and fast-growing company through the challenges and opportunities of this year, balancing his own company's transition to a remote workforce and a surge in demand for Krisp. He's a native of Romania and also a global citizen that's experienced technology leader at companies like Twilio, and he's making his own adjustments to working and leading from a distributed point of view. So today we're going to chat, and thank you so much for being here.

Thank you, Matt. Thanks for the intro. Well, hi everyone, I'm David, CEO and co-founder at Crisp, as Matt mentioned. So happy to join this podcast.

Was there any key biographical detail that was missed that you love to share, or things that people usually don't know?

No, absolutely. I think that was a great, you know, great introduction. I was born in Armenia. I've lived, you know, 10 years in the US. Right now, I'm back in Armenia. Yeah, we are— I'm sure we will go deeper on my background and biography. I'm happy to share as much as needed.

So in 2018, when you started Crisp, What was the thing that you were seeing? Because people weren't on calls or Zoom nearly as much back then. What was the need you were seeing?

Yeah, absolutely. Well, the story behind Krisp is very personal. So I was actually working at Twilio, right? Which is a communication platform and actually SquadQuest. I just figured that SquadQuest is powered by Twilio. But because my family and my friends were in Armenia, I was traveling a lot to Armenia, like at every chance, I guess. So, and because of the time difference, right, like almost 12 hours of difference, I, and when I connect, I was connecting to meetings, it was evening time here. And in evenings, you know, you want to go out, right, with friends and family. So, but that was the time when I needed to join meetings, right, like my daily meetings. I was heading the product security at portfolio. So that means I had many meetings with different teams, and I always wished there was a button I click and, you know, get some privacy. Like, people don't know where I am, they don't know that I'm joining from bars— not necessarily bars, of course, but still.

So almost like a virtual Zoom background but for audio.

Exactly. So I had that need, but I had no idea how to build the technology, and I knew it must be done with machine learning, right? I knew about voice, but not machine learning. But I mean, that's where I met my co-founder and that's how things have started.

I think I first came across Krisp actually on the NVIDIA machine learning blog. Was that like an early— it was very early on. It felt like the company was— I think it was all still free at that time.

Yes. Well, actually Krisp wasn't released at that time yet. Or maybe just launched. And then that blog post was very important for us. We worked on it for a very long time. And that was sort of the first exposure that our company received. And the blog post got actually a lot of visibility. So it was at some point, I believe, like the most shared and visited blog post on NVIDIA developer like AI section. So yeah, it brought us a lot of visibility.

I actually made a mistake early on when I was advocating for Crisp. I, I told people it was from NVIDIA or spun out of NVIDIA. I was so— because the post had seemed, uh, so great, I couldn't imagine that it was a guest post.

Yeah, well, there is a fun story actually behind that. So, uh, yeah, when, when we, we did that post and it was successful, we thought that we need to put that post on Hacker News, and we put a title which sort of implied that it was from NVIDIA so that people open it more. It was like a small hack from us, and it worked out because, you know, Chris, you know, that blog post was in the top 5 of Hacker News that day. Yeah, exciting times.

That might have been where I saw it too. I don't recall exactly, but that would certainly be plausible. So do you— I imagine you're able to kind of turn Krisp on and off on your setup right now. Can you kind of demonstrate how it works?

Yeah, absolutely. So Krisp is on right now, and I'm gonna clap. Okay, I'm clapping right now. I know, like, when I do this with video, it's much more impressive. And now I'm gonna go, and it's a single button, right? When I turn it off, and then I clap. You hear the clap, right? So that's, yeah, I mean, that's the easiest way to demonstrate it. But Krisp is, you know, with COVID and with everything that happened lately, people moving to home, Krisp was very handy with kids at home, you know, with dogs barking at home. So it's It does a great job at removing noise, and happy to actually dig more into like how that works and where Krisp is going.

It reminds me of the Zen koan, like, what's the sound of a single hand clapping? I guess it's like Krisp. One reason I have been advocating for it a lot is that for a good meeting, you don't need video. You could turn video off if it's not working. We're not using any video now, obviously, but if audio doesn't work, the video— the meeting stops. A meeting with video, unless I guess you're really good at American Sign Language or something, you really do need great audio. I find it so distracting when folks have just a ton going on in the background, but I also feel for them because we're all at home. We have kids working from home, all sorts of things. What COVID boost have you all seen?

Yeah, absolutely. Well, we believe that voice is going to continue being a key sort of, you know, key mean, you know, of communication. And it's going to grow actually way bigger than it is now. With COVID we saw like a very large boost in Krisp downloads and usage. I believe it's now like, it's been like 7x growth for Krisp. And because yeah, there was no technology like this, right, in the world. And, you know, when we were just starting, people didn't really— I mean, every person that was showing, was seeing the demo, they, they could relate so much to the app, right, to the, to the, to the problem. But, uh, they didn't really know that that problem exists, right, because we are so used to what we have. So, uh, it took us a while to, uh, you know market is. And we were like, early on, we were having a lot of struggles to explain that there is actually a pain here. But with COVID things have changed because all of a sudden, this has become a big problem because everyone is home and their kids are crying. And there's just a lot of noise coming from the kitchen and everything. So yeah, so this, you know, people gradually have started, you know, spreading the word, and most of the growth has been done by word of mouth, right? So, um, yeah, it's, it, you know, from business perspective, uh, there was a lot of growth, uh, during this time.

How, um, so let's dive a little bit into how Krisp works. So it uses machine learning, and what sort of learning technique does it use?

Let me do like a short intro into noise cancellation in general, right? The state of the art before machine learning, like people usually use like multiple microphones to try to remove noise or cancel noise. Our phones have multiple microphones on them. You know, one of the microphones is close to your mouth, the way you, you know, hold the phone, right? And the other microphones are very far from that microphone, okay, from your mouth.

Like there's one on the back of most phones, right?

Which phones?

There's usually a microphone on the back, like where the camera is.

Yes, yes, exactly, exactly. So it must be as far as possible so that you can sort of, you know, by, you know, subtracting the two audios from each other, I'm just simplifying it, you can isolate the human's voice. And like this technology is deployed on every phone out there, I guess, like more or less expensive phone, right? And that technology also exists on our laptops, but it just doesn't work because your mouth— the person is very far from the laptop. So, uh, it has two problems, right? One problem is that it requires multiple microphones, so it requires specific hardware. The second is it has limitations on how much noise it can remove. Usually it's great with removing like stationary noise, like static noise. But when the noise comes and goes, like clapping, like, you know, barking, right? It's just not possible to adapt to these sort of noises. And then, like, in the last 5 years, as machine learning has started to grow, like, people have started, like, in academia, they started trying machine learning for noise cancellation. And we were very early on in this problem. So when I met my co-founder, my to-be co-founder, and we started talking about this, we knew that we need to solve this with machine learning just by intuition, right? And we started looking at this, you know, what's out there. As a technology company, we were the first to actually implement, design and implement such technology, which purely uses machine learning for this problem. So the way it works is, We have a very large dataset of like background noises and which we had to find from somewhere, right? It was tough to do that. And but we were clever, I think, with that. Like we had, we tried some interesting, we found the right sources for that, right? And then like these are like very different types of noise, like 10,000+ type of noises. And then we also have collected a lot of like clean studio recordings where there is no noise at all. And so we have a lot of such data, and when we mix them together with different sound-to-noise ratios, right, we get pretty much like infinite dataset of noisy speech for which we have the clean speech, right? Because the we use that dataset. So, and then what we do, we have designed this spatial neural net for which during the training we say, well, this is a noisy speech, this is like a clean speech, noisy speech, clean speech, noisy speech, clean. And we do it for, for all these artificially generated noisy speech, you know. And then it starts to learn what's human speech, you know, what's clean speech, what's noise. And then during the inference, like when, when you start using it, even if it sees like noise types that it never saw before, it's able to recognize them and separate it from each other. So this is very simplified explanation of how it works. Obviously, there is a lot of IP. It's, you know, audio is very difficult, it turned out. If we knew where we are, I mean, we were not audio experts. Our team is very strong at math and physics, but we didn't have any experience in audio. And I always say, I always say, If we knew how difficult audio is, like, we would be just scared off and we wouldn't start this, right? And I'm, yeah, we were lucky that we didn't know that because many teams who have prior audio experience, they are still struggling with this. Yeah.

What do you think was it about not knowing audio that allowed you to take a different approach or succeed where others haven't been able to yet?

Yeah, absolutely. So this is sort of the classical approach to the to the audio problems, right, to digital signal processing problems. Like DSP, digital signal processing, like the theory and like everything is there for like 30 years or 3, 4 years it has been out there. And you, if you are like a DSP engineer, audio engineer, like building microphones and speakers, you are trained to think from, from those constraints perspective, right, from this classical theory perspective, from this classical algorithms perspective. If you need to solve something. That's where your brain goes by default. And from our team perspective, like when we started the company, we were 7 people and 6 of them had like, have PhDs in math and physics. I'm the only one who doesn't hold that. So we had, we had a lot of experience in math. We understand, you know, the, the math required for dealing with audio and machine learning, but we didn't know the the existing theories, right? So we— that was easier for us to start doing new things, which was required because when you do the machine learning approach, you don't necessarily need to use a lot of the old stuff that has been developed for 40 years. And I think that was a key difference.

You're also— it sounds like you're compressing the audio a bit and maybe doing some low-pass filter.

We are not. What we do in Krisp is until, well, even today, Krisp is only working with wideband audio. Okay. Up to wideband audio, which means like 16 kHz of sampling rate. That's great for, well, I'm currently using like Bluetooth headset and Bluetooth compresses a ton, right? Yeah, exactly. So it does it by default. But even if I use like a full-band microphone. Krisp today would downsample to wideband before doing the processing. And that's the reason for that is like we have spent a lot of time on optimizing our technology for CPUs, right? There is no such technology running on CPUs. Like people can run those algorithms on GPUs easily, but for CPU, it's very hard to squeeze that. So we have spent a lot of time on doing this. That's one of the reasons why we decided to stick with wideband. In the same time though, like we are like, you know, we stick with wideband as opposed to what? To full band. So downsample to wideband. Yeah, in a week timeframe, we are going to, I believe it's in a week, we are shipping a new version of Krisp that is going to support full band as well. That's been a very long effort. For us to like squeeze these neural nets to understand the higher frequencies of voice as well. And then, but in the same time, be able to run on CPU.

Do you use a GPU if it's available?

No, we are, we only support CPUs.

Why is that?

Well, two reasons, right? We could support NVIDIA GPUs and they're very powerful, right? They're very, it's very easy to run like neural nets inference on these GPUs. And when you do that, the CPU is offloaded. That's great. But, you know, Crisp is used in like enterprise, in by a lot of professional users who don't have GPUs, right? So, uh, our— most of our population, I mean, of users, uh, they, um, they have just CPUs. And there, the GPUs they have, like the Intel GPUs I have on my Mac, is just not capable of running these neural nets. It's just too small for that, right? So we decided to, you know, spend this extra effort, a lot of effort actually, to support everything out there rather than just focus on like one type of hardware deployment, right?

And to also be clear, when you talk about squeezing the neural nets, Krisp all runs locally on your computer, right?

Yeah, absolutely.

Which is awesome. So there's not like the latency of going to the network and the audio data is not being sent anywhere else. It's all happening locally. So like, what does it mean to squeeze it down? Like, are you worried about download size or the runtime or how much CPU it's going to use?

Yeah, like the privacy has been very important for us and we are very, very happy to— that we were able to actually run this locally. We don't think the audio should go to any server, right? Especially in this world, privacy is very important. Yeah, so to explain what it means to squeeze on this for the CPU, so like, as Crisp is a virtual microphone, it sits between the actual microphone, right, and the app. Like, in this case, it's the browser app, you know, running Squadcast. So CRISP is between them and it needs to run its neural net on every audio frame in real time and, and without introducing too much latency. Okay, so that means that it keeps receiving these frames and it needs to not only, you know, not look too much forward, right, in the audio so that it doesn't introduce this artificial latency, but also not spend too much CPU power so that it can keep up with, you know, with our speech. So that's, that's very constraining from engineering perspective, and that means that you need to squeeze and so make your neural net smaller and more efficient and use, I don't know, like the, the right, like, library that fits best for this kind of like mathematical problem, right? So that it runs properly on the target CPU. Does it make sense?

Yeah, and it's only about like 70 megabytes now. Will, when it has this new sort of full-band neural network, will it, will the download get larger?

Yeah, it will, it will get a bit larger, uh, and even, even in the 70 megabytes, we have multiple neural nets, uh, so we have neural nets that work for like, uh, the 8 kHz sampling rate. We have neural net that supports larger. And then like, you know that Krisp— with Krisp, Krisp works bidirectionally, right? So if I have Krisp, I can remove the noise coming from you, uh, and that, that has a different neural net. So there's a lot of engineering actually in this simple app. Uh, we— I, I believe we have like 3 or 4 models like neural net models shipped today. And with, uh, with this, uh, newer version, we are going to have like 6 or 7 models shipped. So a lot. Yeah.

Wow. Yeah, that's actually kind of a fun feature a lot of people don't know about Krisp is that if someone's annoying you with bad audio, you can actually filter them as well so it sounds good, and you would never even know that they have a dog barking in the background or something.

Yep.

Did that, I mean, was that in the original version or did that come up when you were talking to people who weren't using Krisp yet or?

No, that was in the original version and in the original, on the very original version, we like, one of the challenges we had, like we didn't know how to sort of structure Krisp. Like we didn't know whether they will be using more this inbound noise cancellation or outbound. Like what is more important for people, right? And that was such a, like a, interesting question. Like, do you worry more about your noise, or do you, you know, are you willing to pay for canceling your noise, or are you just, you don't care about that and you're willing to pay for others' noise? So like, when we shipped in the very original version, like, inbound was entirely free, and then the outbound was a pro feature. But then we changed. Now it's a freemium product. Crisp comes with 2 hours free every week. And then like if you go to Pro, it becomes unlimited. And the Pro is going to get some more very cool things very soon.

It is such a good deal. What's the latest pricing on it?

Yeah, so right now the pricing is, I think, yeah, it's $40. $40 a year. That's going to change soon because this was a COVID pricing. When COVID started, we started a program with which all the students, you know, university work, universities, right, government workers, hospital workers would get Krisp for free for 6 months. And, and we also dropped the price by 20%, 30%, I'm sorry. And, but like that 6 months, actually we went with that for like 7 months now and we are bringing back the price. It's going to be $60 per year.

Even at $60, when you look at how much money I've had to spend to like make the room quieter, you're basically getting like a full studio and a really great microphone and everything.

Yeah, absolutely. Actually, Like, we plan to keep that price, but we are going to add some very cool things in the near future, like around virtual backgrounds and more, even greater noise cancellation. But yeah, so we are working hard on this.

Cool. I can't wait. I will be a top customer. I did notice, how much latency does it introduce right now?

On algorithmic side? The latency is between 20 and 30 milliseconds. On the app side, our application introduces an additional like 20 to 30 milliseconds. So overall, it's around 60 milliseconds.

Hmm. Yeah, I did do a video where I posted— I think I recorded just using QuickTime video and used Krisp to take out some background noise, and people could tell that my mouth was just a little bit off. So if you had a way to also introduce the delay to the video so it synced up I think that would be pretty nice.

I'm not sure that was Crisp. I mean, usually, like, with Crisp, you wouldn't notice, uh, the difference. I mean, video is doing a lot of things which might contribute to that. Uh, like, we are using Crisp every day with video, obviously, with Zoom. Like, we have never noticed that. And there are a lot of reasons why, uh, you know, you might have latency. But like, I, I mean, everything adds up, obviously., and, uh, this 50 to 60 milliseconds might contribute at the end, right, if there is enough latency. But that's just not enough to be noticeable.

Yeah, on Zoom I've never noticed it. It was only when I was recording this video. So you're right, there might have been something else, maybe in like the HDMI conversion or, or something where it just felt a little bit out of sync.

Yep.

I actually didn't notice it at all, and then some— I started getting some comments about it. I was like, oh, I kind of see it. Kind of like when you use a soundbar with a TV, sometimes it can be just like a little bit off.

Yeah, yeah, it's interesting. Like, we won't see that, like our eyes usually are adopted. But for example, when you use like virtual background in Zoom or Microsoft, you start noticing it. It's there, you start seeing that. And, you know, even when you move, there is a latency.

Yeah. So what's the latest going on with hardware? So for example, I know that old MacBooks had like terrible, terrible built-in microphone, and the latest 16 actually sounds pretty good. You know, I think, um, uh, John Gruber on Daring Fireball posted an audio file just straight off the Mac 16 microphone, and it sounds like they're doing something that's better. So what, what are they changing, and, um, and what do you think about it?

You mean from audio perspective? Yeah, I'm not really sure. Uh, I, I think my Mac is like a high-end, like it's usually using high-end microphone and speaker, although, you know, we are not— I mean, I don't think people are very happy with what that high-end means. Uh, so yeah, I'm, I'm actually using it. I mean, I don't do podcasts, obviously, that's a more important question when you do podcasts. But yeah, for everyday communication, what they have, you know, works great. So, but I'm not like really following what's happening on the platform.

Check it out. I'm kind of curious. And I'm also curious more broadly, how much do you think some of this gets built in by the phone makers, the laptop makers versus needs to be software?

Yeah, I, I have no doubt that this kind of technology is going to be there like in every device in the next like 3 years, uh, 3 to 4 years. Like typically, uh, phone, phone vendors, they don't like to make changes like this kind of changes. They are, they're a bit slow on that because as you can imagine, like they already know how to build these multi-microphone systems, right, on the phone. And, you know, they have everything like the line set up for that. They know what the yield is and everything. So any change there is going to take time, but I have no doubt it's going to happen. Okay. So, and but again, it also depends what is that is going to happen, right? Like noise cancellation. Like even today, Krisp is not perfect, right? So we are spending a lot of time on improving what's out there. And we, like, in the next year, hopefully in the next 6 months, we are going to ship something inside Krisp that is going to be just revolutionary in terms of noise cancellation because it's just going to level, you know, take this to the next level. And I don't think something like that is going to come to hardware. Soon enough, right? It's going to take some more time. But in terms of like when these devices will have like noise cancellation, I have no doubt that, yeah, in the next like 2, 3 years, everyone is going to have some version of noise, like ML-based, machine learning-based noise cancellation on them.

And how do beamforming mics work? You know, I know like some of the new headphones like the Bose and also just like the Facebook Portal or the Alexa devices have these mics that seem to be able to pick you up from all the way across the room.

Yeah. So, the way they work, they have multiple microphones, right, on the device. In, like, when you turn on the device, like, and you start talking to the device, right, it starts to calibrate and it starts to sort of, given that these microphones have are far from each other, they start to understand where the direction is coming from, right? Where the voice direction is coming from. And they start to focus only on that direction. And sort of, again, like using the same technique that I explained, they're trying to ignore anything else that is not coming from that direction. That's beamforming. The problem with that is like when you keep moving around, It needs to recalibrate again and again. And I'm not sure that it's— that technology is able to, you know, fix this problem by just by its own, right? It might be very useful for far field, like for something like Portal or Alexa, which is in a big room, right? And they need to fix, you know, fix the noise problem. But I'm not sure how efficient it is. We are not dealing with these problems. Actually, we are not dealing with far field as well. That's a very, very different problem, although it might be similar, right? But in audio, every problem is so unique. And yeah, we are not dealing with that.

That's interesting. You know, and one more technical question. You had mentioned full band and wideband. How should people think about like their Bluetooth headset versus a USB headset versus other things, and what kind of a type of audio is being captured by the computer that Krisp is receiving?

Yeah, yeah. By the way, I'm not an audio expert. I should, I should say that. We have a lot of audio experts in the company, but I know as much as I know. In terms of Krisp, Krisp doesn't really— it doesn't really matter where the audio is coming from, and that's one of the beauties of this machine learning-based algorithms, right? You can even— like, what we have, you can even run it in the cloud, right? Because it really has no audio— it has no hardware dependency. Let me give the example of inbound, right? So imagine, like, I have Krisp here. You talk and I can cancel the noise coming from your audio. So when you talk, there is so much transformation happening to your voice, right? Starting from microphone, that the microphone has like its own transformations including noise cancellation, and then the browser gets it and then sends it over WebRTC with all the like codecs and everything, and then it receives here, unpacks, and gives to Kris and then Krisp runs its technology on it. So pretty much it doesn't matter where you run, right? You could easily run it in the cloud. So from that perspective, it doesn't matter whether it's a USB, you know, microphone or a Bluetooth microphone, uh, or like just a wired microphone, ordinary microphones. So it doesn't matter. Obviously, like, we had to add the support for all of these, but from audio perspective, it doesn't matter. Like, usually Bluetooth audio has more latency. It's just like there with the Bluetooth transport, right? Like, you might notice that with AirPods, right? If you have AirPods, usually the— I don't know why, but something doesn't work very well with them when it comes to latency. Like, sometimes it's just like too much latency with AirPods, like, and without Krisp, right?

But yeah, I mean, maybe because they have to connect to each other as well as the source device.

Yeah, I mean, the connection is, is one time, right? You connect and then there is a connection, but sometimes it's just the latency adds up. I don't know what they did wrong there, but I'm hope— I hope they will fix it, right? Uh, so yeah, I mean, but with USB, like, it's usually more powerful. Less convenient, obviously. But yeah, I mean, in general, that's the— like, Kris doesn't care really where the audio is coming from. There's one more thing, like, with when you use a Bluetooth headset, and if you just listen to music, it's, you know, it's using its highest frequency, like it uses all the frequencies possible, right? But when you—

Yeah, higher codec, right?

Exactly, yeah. But when you turn on, like when you start a call, right? When you start using the microphone of your headset, it brings everything down to wideband typically, like some—

Wideband is 16 kHz?

Exactly. And the prior versions of these Bluetooth headsets, they would bring your voice to 8 kHz. We know that's not great.

And that's, it's kind of like using fewer colors to paint a picture, right? So it doesn't feel, sound as full or as natural.

Yeah. Yeah. I mean, like every time we use like telephony calls, right? When you call a phone number, most of the world is still using 8 kHz codecs because they just to transmit less data, right? So that's like, I mean, we are used to that, but when you hear like wideband, I mean, sorry, full band and then narrowband, which is 8 kHz, like you will see the difference. You know, it's a big difference. But still, I mean, so if you're using a Bluetooth headset, there is a big, big chance that, you know, it will downsample it to wideband in the calls. And this is because they need to use like, you know, from energy perspective, from processing power perspective, they need, they need to keep it efficient.

It's actually amazing. I recommend folks, if you have an iPhone, try calling a friend who also has an iPhone using FaceTime audio versus just a normal phone call, and it's astounding how much better you can hear them and understand them. It actually makes phone calls pleasant again.

Yeah, yeah, yeah. It's obviously, I mean, it's using VoIP and like the best codecs out there. I mean, telephony still wins, I think, because of the network, right? You know, like the service providers, right? They, like AT&Ts of the world, they dedicate the bandwidth to voice channels, like telephony voice channels, and then the rest is used by everyone else, like the data channel, right? So, So like your VoIP is going to impact it if the signal is not strong enough, but the telephony voice will still be there. So that's, I guess that's more important than the higher frequencies of the voice because you can at least hear each other. But that's going to fix, you know, it will be fixed with 5G. So we will, once 5G is deployed, I'm sure those problems will go away and everyone will switch to VoIP.

Cool. Well, yeah, and actually I think even now they use VoiceOver LTE by default for a lot of, um, by default on like the new iPhones.

But today, I don't know if, I don't know if that technology is even live. I mean, alive. Uh, I don't know if there has been any VoLTE, uh, deployment.

Maybe I'm wrong, but I thought it definitely is on like AT&T and, uh, and Verizon here in the US, but probably not internationally.

Okay.

It does sound a lot better if both sides are on it, but yeah, well, I feel like cell phone calls drop so much anyway. It's funny, when I was a kid, I remember like spending lots of time, hours on the phone, like, I don't know, and then your parents would pick up the phone and like, "Mom, I'm on the phone." But now I feel like people don't do phone calls as much anymore, partially because the quality is so bad. It's very frustrating.

Yeah, I think it depends on which part of the world you are in. Like, as I know, like in Japan and South Korea, like, I guess nobody is using telephony anymore, right? Everyone is on VoIP. And it also depends on the network connection. Like, that's why, like, if 5G is there, I mean, if there is enough bandwidth, like, why would people use the telephony? There is one use case though that I really believe it's going to still thrive is Thrive is phone numbers, right? Phone numbers are, are such a cool, you know, concept. We don't appreciate them that much, but they are, they're, I mean, they are the most deployed, known, understood, like, handles that we have, right? Uh, and, you know, everyone can reach out to you, although with spam, of course, with the problem of spam. But I think that that technology is not going to be go away. I thought about that a very long time. I, and I think it's going to stay around. I think there is a lot of things that you can build on top of phone numbers and it's going to thrive.

I can tell some of your Twilio days coming through.

I know, I know, it's definitely coming from there. Yeah.

Well, just, I really love that we're able to do a deep dive in Krisp. I can't wait to see what you're launching in a week or two. So I'm looking forward to the update. I'm going to do a ton of audio tests and record things with different mics and try it out. So thank you so much. You, you're also running a company, and I know that y'all were mostly in person, I think, in Armenia before. How have you adapted, and how's it been, and what are you planning to do once we can be safely in offices again?

Yeah, absolutely. So our company is distributed between US and Armenia, and we have a team member in Germany as well. So we are distributed. We have a big presence in Armenia I spend a lot of time in Armenia and my co-founder as well. So, um, like before COVID uh, we were spending a lot of time in the office, in Armenia office, although everyone else was remote. Uh, and after COVID, obviously we are, we are working, uh, remotely. And right, right after I think COVID started, uh, I guess like in two, like two weeks in we decided that Crisp must become a remote-first company. And it's not because of COVID we always had that idea, right? Because, you know, it just makes so much sense, especially for us now that, I mean, we are focused on building, you know, a tool that helps remote folks. So we always had that idea, but COVID was just, you know, catalyzed that. And, and yeah, so like We, I think at the end we are going to become a hybrid company, right? Like, because we have the office and a lot of people actually enjoy being in the office a couple days a week and they might not have the right setup in the house, you know, they might have kids. And so, and just like, like I personally like to come sometimes to office because that's how my brain works, right? It needs that environment. But in the end of the day, like 80%, 85% of our workforce today just doesn't come to office because they enjoy working from home. And actually, like one fun fact, we also have a program called Work from Forest. So every week, every week we have like 10, 15 people taken to some nice place, you know, outside. Countryside, and they have internet power and they do hikes and they actually work as well. Yeah. And it's very efficient and productive work happening.

Yeah.

I'm in love with this program.

So this is happening right now. You just are able to do it in a distance fashion.

Yeah.

Work from forest. Yeah. Well, that's a new one for us. I haven't heard that one before. I think it's a great idea.

Everyone, you know, it's quite popular now in our office and it's going— I think it's going to evolve. Like, one problem with that is the weather, right? Whether the weather is good or not. And but I think we will find a solution for that as well. Like, it's a great way for people to still gather together. Right. And, you know, see them, see each other in person, which— and that's a very important part of building culture and relationships., but they don't have to come and sit in the office for that. Yeah. And we are determined to continue growing as a remote-first global company. So we are very excited about that.

Awesome. Well, I can't wait to see what's next. David, thank you so much for coming on. If you're listening, please get Krisp.ai as soon as possible so your calls sound better. I guess if you get in soon, you can get the $40 pricing, but it's a good deal even at 360. And, um, and yeah, as you go distributed, make sure to check out the other episodes of the Distributed podcast. There might be some good tips for you or your managers there. So thank you again for coming on.

Thanks a lot, Matt.

All right, you have been listening to Distributed with Matt Mullenweg. Please subscribe or tell your friends or rate it, and we'll keep doing this. I'll see you next time. Bye-bye.