Latest / The Adults In The AI Room / Artificial Voice Intelligence - Maxim Serebryakov - Earley AI Podcast with Seth Earley & Chris Featherstone - Episode # 029

Transcript

Artificial Voice Intelligence - Maxim Serebryakov - Earley AI Podcast with Seth Earley & Chris Featherstone - Episode # 029 · The Adults In The AI Room

• 0:01
  SPEAKER_01: This is the Early AI Podcast, hosted by Seth Early and Chris Featherstone.
• 0:06
  SPEAKER_01: Join us as we delve deep into the passions, expertise, and experiences of thought leaders and practitioners to talk about what's possible in artificial intelligence.
• 0:15
  SPEAKER_01: Early AI Podcast is sponsored by Early Information Science, your digital transformation journey with the design and deployment innovative technology solutions.
• 0:27
  SPEAKER_04: Welcome to today's podcast.
• 0:29
  SPEAKER_04: I am Seth Urwey, and I'm Chris Featherstone.
• 0:32
  SPEAKER_00: It's good to be with you, my friend.
• 0:34
  SPEAKER_04: Yeah, nice to see you.
• 0:35
  SPEAKER_04: Our guest uh today is passionate about using AI to give people a choice in how their voice sounds to others.
• 0:42
  SPEAKER_04: Born in New York, raised in Russia, and educated at Sanford Creek.
• 0:46
  SPEAKER_04: Please welcome co-founder and CEO of Samus, Max Sarabriakov.
• 0:52
  SPEAKER_00: You know, he's passionate about helping us understand how to pronounce things correctly, right?
• 0:59
  SPEAKER_02: Well, the thing you realize, Chris, is um there's really no way of pronouncing anything correctly.
• 1:05
  SPEAKER_02: There's so many different deviations and variations.
• 1:07
  SPEAKER_02: So yeah, it's it's a pleasure to be here, guys.
• 1:09
  SPEAKER_02: Thank you for inviting me and um yeah, looking forward to the chat.
• 1:13
  SPEAKER_04: Yeah.
• 1:13
  SPEAKER_04: So tell us a little bit about your background.
• 1:15
  SPEAKER_04: Tell us uh the world according to Max and you know what's your what's your uh journey been and how did you end up where you are?
• 1:22
  SPEAKER_04: So like to hear the background.
• 1:25
  SPEAKER_02: Yeah, totally.
• 1:26
  SPEAKER_02: So I'm Alex.
• 1:27
  SPEAKER_02: Um I was originally born in New York.
• 1:31
  SPEAKER_02: Uh didn't live for for quite some time in New York.
• 1:34
  SPEAKER_02: I I basically was born there, lived a few months, and then uh moved out.
• 1:38
  SPEAKER_02: Um lived on for most of my life in Russia.
• 1:42
  SPEAKER_02: Um I grew up there, my family is in Russia.
• 1:45
  SPEAKER_02: So I as you can imagine, I've had quite a few very unusual perspectives when it comes to uh coming to the US, um accents, um being an immigrant.
• 1:55
  SPEAKER_02: Uh and that's really sort of the the brewing grounds for the creation of my company, Sonus.
• 2:01
  SPEAKER_04: So you when did you move back to the US and what brought you to the US?
• 2:04
  SPEAKER_02: Yeah, great question.
• 2:05
  SPEAKER_02: So I I got into uh Stanford uh from high school uh when I was 18.
• 2:11
  SPEAKER_02: Uh came to Stanford.
• 2:13
  SPEAKER_02: It's uh such a beautiful place.
• 2:15
  SPEAKER_02: Um I've done so many different uh trips across the US with my family, and um it's it's incredible out here.
• 2:22
  SPEAKER_02: Um yeah, you you know, I um um I studied at Stanford and part of uh my studies, I spent a lot of time um looking into artificial intelligence, studying artificial intelligence.
• 2:36
  SPEAKER_02: Um and uh through these experiences, you know, I ended up uh building saunas and here we are today.
• 2:43
  SPEAKER_00: But originally, originally you weren't looking at it from language and speech, because you had you had dabbled into a bunch of different other things.
• 2:51
  SPEAKER_00: Like I think we we had talked at one point that you were looking at quant style mathematics and you were looking at you know all sorts of interesting statistical anomalies, all that kind of stuff.
• 3:02
  SPEAKER_00: I was working on language and speech, and I had actually tried this scenario and this use case centered around we had had um a group down in the south and they were trying, and they were a trucking company, and they had their call center in India, right?
• 3:19
  SPEAKER_00: And so as you can imagine, there was um really, really low, you know, like just abilities to communicate correctly because we had, you know, the the the Hindi English speakers and then these southern truckers, and they were trying to understand frustrations, all that kind of stuff, right?
• 3:38
  SPEAKER_00: Just in terms of just broad communication.
• 3:40
  SPEAKER_00: And so what we had look, you know, what they were asking for was there a way to take the speech recognition in and take you know what was being said and automatically um normalize for all intents and purposes, the voice so they sounded the same coming in and out.
• 3:59
  SPEAKER_00: And that was just all on just sound, right?
• 4:01
  SPEAKER_00: Just like, hey, I want my voice to sound like theirs and vice versa, even though they're speaking.
• 4:06
  SPEAKER_00: Well, as you can imagine, you know, it sounds simple to be able to do, but there's a lot of complexities in that because part of that too was the latency of going from um real-time voice communications and going up to the cloud and back.
• 4:24
  SPEAKER_00: And so essentially I failed.
• 4:26
  SPEAKER_00: I failed at being able to provide this kind of thing because the latency was too great.
• 4:31
  SPEAKER_00: We you couldn't have more than you know, even two or three seconds of lag within that and still try to handle a conversation.
• 4:39
  SPEAKER_00: We've been on long distance phone calls before, right?
• 4:42
  SPEAKER_00: Where it's like, hey Seth, how you doing?
• 4:45
  SPEAKER_00: I'm doing really good.
• 4:47
  SPEAKER_00: You know, I mean, it just it creates you know this really thing.
• 4:50
  SPEAKER_00: So I failed at it, but it taught me a a lot about you know what need what the the world needs in terms of, and I just didn't have, you know, I mean, I it wasn't part of my core use cases.
• 5:01
  SPEAKER_00: So I was like, great, um, that was an interesting, you know, science science project.
• 5:06
  SPEAKER_00: Put it off to the side and we'll go focus on you know other things.
• 5:09
  SPEAKER_00: Well, fast forward to you know, like three or four years, and I had a friend that called me and he's like, hey, uh, I just ran across this really, really slick company, and he's not technical.
• 5:20
  SPEAKER_00: And he's like, they do this.
• 5:22
  SPEAKER_00: And I'm like, actually, I've already done that before, but I couldn't get to work.
• 5:25
  SPEAKER_00: So I'd be really interested in how they got it to work.
• 5:28
  SPEAKER_00: And he's like, Well, I'm gonna put you in touch with Max.
• 5:32
  SPEAKER_00: And the our uh, you know, our beautiful relationship started there, right?
• 5:37
  SPEAKER_00: We were able to go in it, but anyway, so I digress.
• 5:40
  SPEAKER_00: But Max, I do I do want you to get back to like a little bit of how you got to this piece because it's a really neat story.
• 5:47
  SPEAKER_00: What you were focused on, because you weren't focused on speech and language before.
• 5:50
  SPEAKER_02: But what's interesting about the Stanford computer science education and uh artificial intelligence education is you basically take a bunch of different classes um and you uh learn really the fundamentals in all of these different categories of machine learning, like computer vision, uh natural language processing, spoken language processing, you learn techniques, you know, reinforcement learning techniques, and um a lot of these state-of-the-art architectures as well.
• 6:18
  SPEAKER_02: So you really get a good broad perspective on how things work.
• 6:23
  SPEAKER_02: Um, you get a good education on like probability theories, stochastic processes.
• 6:28
  SPEAKER_02: Um but really your specialization starts whenever you you know you leave college, right?
• 6:35
  SPEAKER_02: And um the areas of token.
• 6:38
  SPEAKER_02: So uh what pulled us into building Sonus was really a fascinating incident.
• 6:44
  SPEAKER_02: Um uh that happened to one of my good friends, uh this guy named Raul.
• 6:48
  SPEAKER_02: Uh Raul was a Stanford systems engineer uh at the time.
• 6:52
  SPEAKER_02: He was a very, very smart guy.
• 6:54
  SPEAKER_02: And it just so happened that a few years back, Raul's home country, Nicaragua, was going through a very tumultuous time period.
• 6:59
  SPEAKER_02: They were going through a revolution, and uh the demonstrations uh from uh this left his family in pretty uncomfortable financial situations, right?
• 7:08
  SPEAKER_02: And uh he had to uh sadly take a leave of absence uh from his college degree to support his family financially, uh, which involved him just trying to find a job.
• 7:19
  SPEAKER_02: Um he was an immigrant uh and he came back uh to his home country, Nicaragua, and he started applying to software engineering positions uh day in and day out.
• 7:29
  SPEAKER_02: What was interesting is um we all know Stanford.
• 7:33
  SPEAKER_02: It's a great school, we all know these Ivy League colleges, but in a place like Nicaragua or you know, places outside of the U.S., they don't actually know what Stanford is.
• 7:41
  SPEAKER_02: They don't know it's a great school, that uh he's technical, he's talented.
• 7:45
  SPEAKER_02: When they see his resume, all they see is hey, this is a person that um uh that uh um you know never matriculated, they don't have their bachelor's degree, they only have a high school diploma, and they have very high English proficiency skills.
• 7:59
  SPEAKER_02: So for a high-paying software engineering job, you really can't, you know, get this position.
• 8:04
  SPEAKER_02: So then his horizons expanded and he started applying to less conventional positions that are less fitting for him.
• 8:10
  SPEAKER_02: He started uh applying to you know call center jobs.
• 8:13
  SPEAKER_02: Uh interestingly enough, he ended up getting a call center job working in a call center that answers customer queries associated with the broken computers, uh, Stanford systems engineer, right?
• 8:22
  SPEAKER_02: A guy that could disassemble, disassemble a computer and knows the ins and outs of an operating system, basically just day in and day out answering these very repetitive queries.
• 8:31
  SPEAKER_02: And what's fascinating is you think that a profile a profile such as Raul's would outperform everyone within the contact center.
• 8:39
  SPEAKER_02: Um, but the truth is actually quite distinct.
• 8:41
  SPEAKER_02: He was underperforming.
• 8:43
  SPEAKER_02: And the reason why he was underperforming wasn't because of a lack of knowledge, wasn't because of him not knowing what to say, it's because of how he communicated with his customer base.
• 8:51
  SPEAKER_02: Uh, he had a very thick South American accent, which led to people treating him differently.
• 8:56
  SPEAKER_02: And what do I mean by different, right?
• 8:57
  SPEAKER_02: From one end of the spectrum, these were full-blown racial slurs, to the other end of the spectrum of people just questioning his qualifications solely on the basis of how he sounds.
• 9:07
  SPEAKER_02: And after learning about this, we started exploring this idea a little bit more of, you know, what if we were able to transfer Raul's accent so that um, you know, accent is no longer a function of you know his success, right?
• 9:21
  SPEAKER_02: Or uh a function of the overall equation.
• 9:24
  SPEAKER_02: Limited limiter in the song.
• 9:25
  SPEAKER_02: Exactly.
• 9:26
  SPEAKER_02: It's no longer a limiter at all.
• 9:28
  SPEAKER_02: And we started building these algorithms.
• 9:31
  SPEAKER_02: Um and you know, we started off building off some of these uh voice conversion uh research um uh articles that were outward.
• 9:39
  SPEAKER_02: So we've implemented some papers and uh thinking that voice conversion could actually probably transfer someone's uh uh accent.
• 9:47
  SPEAKER_02: You know, but quickly we saw that you could convert a person's uh voice to sound like, say, I don't know, a different gender or some superhero like Batman.
• 9:56
  SPEAKER_02: But if the input speech is Indian accented or Russian accented, you're gonna have a Russian-accented Batman as an output, uh, which is funny to imagine, but it shows the limitations of modern-day voice conversion research in that you're not just modulating the pitch and tone, you're actually changing the underlying phonemes that are present.
• 10:14
  SPEAKER_02: Um what was a phoneme?
• 10:16
  SPEAKER_02: Phoneme is like the most granular level of speech that you could get, right?
• 10:20
  SPEAKER_02: In images, you're working with RGB pixels, right?
• 10:22
  SPEAKER_02: Pixels is the most granular uh for speech, it's it's phonemes.
• 10:26
  SPEAKER_02: And uh, yeah, so you start exploring this.
• 10:29
  SPEAKER_02: We tried also, Chris, a little bit later, implementing some speech to text, text-to-speech.
• 10:34
  SPEAKER_02: Uh I remember when we grabbed dinner for the first time, we had a good chat about that.
• 10:38
  SPEAKER_02: And I could tell you spent a lot of time thinking about this.
• 10:41
  SPEAKER_02: And our biggest limit shows the side the same, right?
• 10:44
  SPEAKER_02: Latency.
• 10:45
  SPEAKER_02: Um, there are a lot of limitations that come from prosody.
• 10:49
  SPEAKER_02: Um, you know, text is a very lossy form of data.
• 10:52
  SPEAKER_02: Um, you go from uh high amount of data to pretty much very flat amount of data, and then from that you synthesize.
• 10:59
  SPEAKER_02: So there are all sorts of limitations that come from that.
• 11:02
  SPEAKER_02: But yeah, we ended up uh um building an algorithm that really doesn't exist in the research world.
• 11:07
  SPEAKER_02: It's very innovative, it works on the edge, uh, works on the client, um, minimal latency, and it's very streamlined um efficient.
• 11:17
  SPEAKER_00: I mean, though the one thing too that we discussed was that this is not just a problem that can be solved by a general speech model.
• 11:26
  SPEAKER_00: Like a like doing speech recognition is one thing, but this is this is so much more complex, centered around the sound models that need to go into it and like you said, how to get to the the um the most basic constructs of what makes up a voice, which is sound, right?
• 11:45
  SPEAKER_00: So it can't be, you know, just you know, because there's I know there's a lot of people out there like, well, hell, I'll just do this with some speech rack and like you said, some text-to-speech, and then I'm good.
• 11:55
  SPEAKER_00: No, you can't do that.
• 11:57
  SPEAKER_00: You know, it doesn't translate.
• 11:59
  SPEAKER_00: And so especially with, you know, when you start talking, like you said, you get into tonality, sound, you get into inflection, sound, you get into speed, sound, right?
• 12:10
  SPEAKER_00: There's all of those, you know, key pieces and also the non-verbal things that are going on as well that you have to and then how to conjugate and everything else.
• 12:19
  SPEAKER_00: So, anyway, yeah, yeah, it's a fascinating problem.
• 12:23
  SPEAKER_02: Chris, you you know, you you you nailed it, right?
• 12:25
  SPEAKER_02: Um, like uh a lot of these um how how do you uh map properly disfluencies?
• 12:31
  SPEAKER_02: You know, what makes speech natural are the ums and the uh uhs and the elongated photons.
• 12:38
  SPEAKER_02: Um so um it's a really really fascinating research problem.
• 12:43
  SPEAKER_02: And for us, uh the applications for this technology are very broad.
• 12:48
  SPEAKER_02: We could bring it into education, we could bring it into um you know enterprise communications, into medicine, and all of these different spaces that are very relevant where um uh accent could be uh something that limits the speaker.
• 13:04
  SPEAKER_02: Um and uh um really we chose uh deployments initially with contact centers and enterprises uh because we saw that um the pain point is most clear initially um with them, uh with these deployment verticals.
• 13:21
  SPEAKER_02: And on top of that, also um speech is very structured, right?
• 13:26
  SPEAKER_02: Like you're not working with any whispering speech or yelling speech or crying speech, any and you working largely with monotone or slightly enthusiastic tonalities.
• 13:36
  SPEAKER_00: I mean, it it was probably the the the baseline to do English, right?
• 13:40
  SPEAKER_00: I mean that's a that's an easy one.
• 13:41
  SPEAKER_00: I mean, the majority of the world speaks in English, especially in business, so I get it.
• 13:45
  SPEAKER_00: But um you have to solve for way more than just English.
• 13:48
  SPEAKER_00: You know, what you know, what is Sonus doing now in terms of multi-languages?
• 13:53
  SPEAKER_02: Chris, you say that, but funny enough, we we've actually done deployments for other languages.
• 13:57
  SPEAKER_02: Um uh and interestingly enough, the learnings with us is the because we take our analysis on a phony level, uh, phonemes are pretty similar across different languages, right?
• 14:08
  SPEAKER_02: So um it ends up being a model that's robust enough to account for many different languages.
• 14:14
  SPEAKER_00: That's awesome.
• 14:16
  SPEAKER_04: Yeah, fascinating.
• 14:17
  SPEAKER_04: So how long have you had the business and then uh what kinds of uh clients do you targeted out?
• 14:26
  SPEAKER_02: Yeah, we work with some of the largest contact centers in the world.
• 14:30
  SPEAKER_02: Um you know, of uh you know the top call centers that you know we work with the majority of them.
• 14:38
  SPEAKER_02: Um we work with top enterprises as well, um, for whom uh customer service interactions are pretty key.
• 14:45
  SPEAKER_02: Um mostly just deploy on on voice-based call centers, that's our priority, um, just given the nature of our tech.
• 14:52
  SPEAKER_02: Uh prioritize mostly most of our deployments are in Indian Philippines.
• 14:56
  SPEAKER_02: Uh, we have expanded out to uh deployments in Pakistan and a few places in Europe as well.
• 15:02
  SPEAKER_02: Uh so we're very actively growing.
• 15:04
  SPEAKER_02: Oh, and also um we have some deployments in South and Central America.
• 15:09
  SPEAKER_02: And these spread across uh enterprises and call centers.
• 15:14
  SPEAKER_00: I'd love to get a take too, Max, on you know, what the you know, it's the intended consequences are really, really interesting, right?
• 15:23
  SPEAKER_00: In terms of providing, you know, more value, better communication for people.
• 15:28
  SPEAKER_00: And I think that's the best part is that it it really the the core of what you're doing is is communication is a core element to life.
• 15:37
  SPEAKER_00: And if you can better that in some way, form, or fashion, then guess what?
• 15:41
  SPEAKER_00: You you better the quality of life, right?
• 15:44
  SPEAKER_00: And that's what I love about the um, you know, the intended benefit.
• 15:48
  SPEAKER_00: What are some of the unintended consequences you've seen from it so far?
• 15:54
  SPEAKER_02: You know, from what I've seen, like what's interesting is in our deployments, um one of the most heartbreaking things that we've had to encounter is um uh people going through accent training.
• 16:08
  SPEAKER_02: Right.
• 16:08
  SPEAKER_02: So say a person comes into the contact center uh and uh they get rated a certain level of fluency, right?
• 16:15
  SPEAKER_02: They have different um levels of fluency, and they have this term called mother tongue influence as well.
• 16:20
  SPEAKER_02: And uh say they want to get upgraded uh from one level of fluency to another level of fluency, the the way they oftentimes do that is um actually by uh going through um accent training.
• 16:32
  SPEAKER_02: Um and all sorts of these different enterprises spend massive amounts of money um uh putting agents through accent training to modify the way they speak uh for them to fit in uh better and uh to communicate um um with their customers uh in the desired accent.
• 16:50
  SPEAKER_02: So with us, actually, you remove the need for accent training completely, right?
• 16:55
  SPEAKER_02: A lot of call centers reject agents also in the recruitment type pipeline um solely on the basis of accent two, right?
• 17:02
  SPEAKER_02: That's no longer a function.
• 17:04
  SPEAKER_02: You no longer reject candidates solely on on how they sound.
• 17:08
  SPEAKER_02: There are all sorts of these unintended consequences because one thing that we realized is this is a really important problem for our customers, and they really, really care about this um a lot because at the end of the day, um you know they care about their employees and they want to make sure that they're they're happy.
• 17:25
  SPEAKER_02: And this just removes one thing uh for anyone to think about.
• 17:28
  SPEAKER_02: And in in general, when people come uh taking a more broad perspective, like when it comes to vision, um you know, there's similar parallels between Sonus and uh the the way things work when when it comes to um uh like Instagram.
• 17:45
  SPEAKER_02: Like on Instagram, take a picture of yourself, you could make yourself look skinny, you could make yourself look fat.
• 17:50
  SPEAKER_02: You have uh complete control over the way you represent yourself.
• 17:52
  SPEAKER_02: You could change your eye color, you could change your hair color, whatever you want, you can make it happen.
• 17:57
  SPEAKER_02: Um but Instagram, it's just a picture, right?
• 18:01
  SPEAKER_02: It's very one-dimensional.
• 18:02
  SPEAKER_02: Um, Chris, what you mentioned earlier is very true.
• 18:05
  SPEAKER_02: Like what matters is really your voice, right?
• 18:07
  SPEAKER_02: Who you are, the content of your character is portrayed through that, um, not just your appearance.
• 18:12
  SPEAKER_02: And why is it that you have a level of control like that when it comes to your digital appearance and photographs?
• 18:18
  SPEAKER_02: You don't have a similar level of control when it comes to your voice.
• 18:20
  SPEAKER_02: And we're really excited about being that medium, the platform that empowers people to um feel comfortable with their voice and uh control the way they speak and not have to change the way they speak to fit in.
• 18:33
  SPEAKER_00: That's fascinating.
• 18:35
  SPEAKER_04: What are what are some of the challenges with uh doing a deployment uh in an enterprise?
• 18:39
  SPEAKER_04: And imagine you need to train on specific uh types of I mean, I'm sure there's there's generalized models and specific models that you have to find to.
• 18:51
  SPEAKER_04: What what are some of the challenges to deployment?
• 18:53
  SPEAKER_04: What what does an organization need to do to be ready to use it like yours?
• 18:58
  SPEAKER_02: Yeah, you know, great question.
• 18:59
  SPEAKER_02: For us, uh, from the model perspective, I don't think there are as many issues.
• 19:04
  SPEAKER_02: Um things are mostly fine.
• 19:07
  SPEAKER_02: Um the reason why is because um we built a model that's very generalizable, like it's able to generalize across different dialects and accents exceptionally well.
• 19:18
  SPEAKER_02: And we have a very large data internal to Sonus that um prioritizes it.
• 19:23
  SPEAKER_02: Um and a good distribution of different dialects.
• 19:27
  SPEAKER_02: Uh but really uh for us, one of the challenges is just working with customer tech stacks, right?
• 19:33
  SPEAKER_02: Different customers have different OS systems, different failed versions of OS systems, uh, different hardware, different chips, right?
• 19:42
  SPEAKER_02: Some have Intel chips, some have AMD chips, uh, some have two core processing computers, some have four.
• 19:49
  SPEAKER_02: You know, accounting for each and every variation of hardware is definitely no easy task, but it's a necessary task when deploying on the client.
• 19:57
  SPEAKER_02: Um, you know, some folks have Linux operating systems, Windows OS, Google OS, good stuff.
• 20:04
  SPEAKER_00: I mean, I can think about this too in terms of Seth, your question around, you know, some of the barriers to entry for organizations, because part of this is what as well from what I've seen is that it creates if you can, if you can, you know, help figure out what the accent is, then you can also generate more accuracy to refine things like search capabilities or you know, or you know, like because part of this too is I know in your mind you're probably thinking about how in the hell do I put together an information architecture centered around accents?
• 20:36
  SPEAKER_00: It just does, it's not, it's not there.
• 20:38
  SPEAKER_00: But at the same time, if we're pulling back, you know, the information architecture is all the data that's being generated from all of these things, then I get more accurate transcripts, I get more accurate documentation, you know, and probably better abilities to understand sentiment and scoring, blah, blah, blah, blah.
• 20:57
  SPEAKER_00: Right?
• 20:57
  SPEAKER_02: Yeah, like so true.
• 20:59
  SPEAKER_02: Like one of the biggest limitations modern-day speech to text engines is how robust are they towards different edge cases, right?
• 21:06
  SPEAKER_02: Um, there's a lot of bias in speech to text that lead to it being synthesis, it it recognizing a certain gender speech better, or certain uh pitch distribution better, or a certain, I don't know, accent better.
• 21:18
  SPEAKER_02: And accents are really big one.
• 21:20
  SPEAKER_02: Um, and a lot of these like massive big tech titans are really fighting for a couple percentage improvements in word error rates for these speech text engines.
• 21:30
  SPEAKER_02: And if you get a certain level of improvement, you know, you're popping champagne and you're celebrating.
• 21:34
  SPEAKER_02: Um, and this improvement is really just built off of robustness to these edge cases that I just mentioned.
• 21:40
  SPEAKER_02: And one cool little thing that Sonus provides is it normalizes a lot of the input audio, normalizes it to sound like you know, whatever your audio distribution is able to understand best.
• 21:52
  SPEAKER_02: And oftentimes, because the way speech text engines are trained nowadays, they're mostly trained on uh standard American English speech.
• 21:59
  SPEAKER_02: Uh So if you're able to um build a algorithm and use Sonos as a front end for for that speech to text algorithm.
• 22:07
  SPEAKER_02: That's pre-processing the pipeline.
• 22:09
  SPEAKER_02: Exactly.
• 22:09
  SPEAKER_02: It also improves some.
• 22:17
  SPEAKER_02: Yeah.
• 22:17
  SPEAKER_02: It's an interesting idea.
• 22:19
  SPEAKER_02: Um very interesting idea.
• 22:21
  SPEAKER_02: Um and it also like democratizes speech to text a little bit, right?
• 22:26
  SPEAKER_02: Like right now, any of us could go on GitHub, download 100,000 hours of speech data, and um use that speech data to train a speech-to-text engine, and you get pretty solid results from that.
• 22:42
  SPEAKER_02: But it won't be robust to different accents, right?
• 22:45
  SPEAKER_02: Because um these Titans have their own, these massive tech companies have their own data sets off of which they train, so you probably won't have access to them.
• 22:54
  SPEAKER_02: Um but you can have access to these open source ones and get a good enough speech-to-text engine that you can use internally.
• 22:59
  SPEAKER_02: Speech to text can be really expensive for companies too.
• 23:02
  SPEAKER_02: So this could really be for companies that are okay with a certain level of accuracy, this could really be a game changer.
• 23:08
  SPEAKER_00: Um hardest accents, which are they?
• 23:13
  SPEAKER_02: That's a good question.
• 23:14
  SPEAKER_02: Hardest accents for what specifically though?
• 23:16
  SPEAKER_02: Or just in terms of just yeah, just what what's been your what's been the hardest one of the favorite things about my job is really how much you learn about different cultures, different accents, and uh the linguistics that come with these different um places on earth.
• 23:30
  SPEAKER_02: And uh it's fascinating.
• 23:33
  SPEAKER_02: Uh an interesting one is like um an interesting accent is uh the Japanese accent, I would say.
• 23:40
  SPEAKER_02: Um I wouldn't say it's it's hard, uh, but it's it's very unusual relative to any other languages that are out there.
• 23:48
  SPEAKER_02: Um why uh that's largely because first, sometimes what happens in the Japanese language is they flip phonemes.
• 23:57
  SPEAKER_02: Um sometimes uh L becomes the R, right, when uh Japanese accented individual speaks in English.
• 24:05
  SPEAKER_02: Um that sometimes happens.
• 24:07
  SPEAKER_02: Uh another thing that that's very common in the Japanese language is oftentimes they have consonant, vowel, consonant, and vowel-based combinations, right?
• 24:17
  SPEAKER_02: If you look at um you know, cities in Japan, that's usually the combination.
• 24:22
  SPEAKER_02: And so words that have a lot of consonants together are a little bit difficult for um Japanese accented individuals to pronounce.
• 24:30
  SPEAKER_02: Um for example, like the word McDonald's.
• 24:34
  SPEAKER_02: Like in Japan, they add in extra vowels next to the consonants.
• 24:37
  SPEAKER_02: So sometimes the way they say McDonald's is no, it's maku donarudo.
• 24:44
  SPEAKER_02: Maku Maku Donarudo.
• 24:46
  SPEAKER_02: It's a very interesting one.
• 24:48
  SPEAKER_02: But we we get a lot of traffic from Japan though, massive.
• 24:51
  SPEAKER_02: And we're definitely going to expand out there.
• 24:52
  SPEAKER_02: It's just uh there's some of these very interesting research problems, also.
• 24:57
  SPEAKER_04: I just wanted to pause for a moment, uh, we're about halfway through, and remind uh folks that were speaking with uh speech conversion expert, uh Max Sarabriakov.
• 25:08
  SPEAKER_04: Get it a little bit better that day.
• 25:12
  SPEAKER_00: So what's interesting is is um I've I I've had to give this uh this speech in internally around um how speech works and things like that.
• 25:22
  SPEAKER_00: And part of it heartbreaking enough is that um, you know, because you always want to pull data as to how this works and you know, and where speech centers are and how you learn and everything.
• 25:34
  SPEAKER_00: And, you know, come to find out in 2022 in the US, 21% of adults are illiterate.
• 25:41
  SPEAKER_00: Now you you take out the factors of let's say, you know, 10, 10 to 15% of those are non, you know, non-native US people.
• 25:51
  SPEAKER_00: So that means that there's a huge generation of folks that are illiterate.
• 25:55
  SPEAKER_00: And then there's another 60 or 70% of them are sixth grade and below, right?
• 26:00
  SPEAKER_00: Literacy.
• 26:01
  SPEAKER_00: So I'm always fascinated by these types of things because then you start to look at it from the perspective of, well, then where do you learn, you know, in terms of, you know, because we always learn speech before we learn reading.
• 26:15
  SPEAKER_00: And one is in the language processing part of the brain, and the other one is in the visual processing part of the brain.
• 26:21
  SPEAKER_00: And so it's been super interesting, especially with these types of technologies now that are coming out.
• 26:27
  SPEAKER_00: Um, you know, there's definitely, and which is interesting because when you read to your kids, you're doing exactly that.
• 26:33
  SPEAKER_00: You're showing the visual and the speech at the same time.
• 26:37
  SPEAKER_00: So they're getting both of it, you know, so it's almost accelerated reading, like subtitles.
• 26:41
  SPEAKER_00: You know what I mean?
• 26:42
  SPEAKER_00: So then when you get these types of scenarios coming out, um, it's all processed in the language part of the of the brain.
• 26:49
  SPEAKER_00: So, you know, to your point, Max, you know, we should see start people start to really understand quicker and better.
• 26:56
  SPEAKER_00: But you bring up the most valid point, which is, you know, part of what's not there is the influence of culture, the influence of of you know how sentence structures are put together by language, um, really like Cyrillic versus you know, uh kanji versus katakana versus English versus Spanish.
• 27:18
  SPEAKER_00: I mean, it's you know, and now you start to think about the millions and billions of combinations that you guys are actually solving for, and it's absolutely phenomenal and fascinating, you know.
• 27:30
  SPEAKER_02: Yeah, and uh completely agree.
• 27:32
  SPEAKER_02: Thank you very much.
• 27:33
  SPEAKER_02: And you know, we're still early in the journey.
• 27:37
  SPEAKER_02: There's still a lot of different other options and edge cases to to account for.
• 27:41
  SPEAKER_02: Um like uh you know, duration-based accents are also very interesting ones where you know it's more elongated speech, shortened speech, uh you know, all sorts of these different deviations and like UK-based accents.
• 28:00
  SPEAKER_00: What's an accent, what's an example of a of a duration-based accent?
• 28:03
  SPEAKER_02: Like um Japanese, for example, they're adding additional phonemes that increase the the duration.
• 28:10
  SPEAKER_02: So you could map these phonemes to synthesize towards uh a target phoneme, right?
• 28:14
  SPEAKER_02: Uh, but if it's uh of deviated duration, it becomes a little bit more difficult of a problem.
• 28:21
  SPEAKER_02: Um in certain uh accents within India, I mean India has a massive number of accents, um certain accents in India they um speak a little bit more quickly uh than other accents.
• 28:33
  SPEAKER_02: Um and just the sheer phonetic diversity that's present in India.
• 28:38
  SPEAKER_02: Like if you look at India overall, in the northeastern India, they have uh Manipuri and uh Assamese-based accents, um, which sound a little bit it's hard to describe, maybe like Australian.
• 28:51
  SPEAKER_02: Um it's hard to compare accents because each accent's so different uh different, but that's the sim most similar one.
• 28:57
  SPEAKER_02: And then you go towards um the southern portions of India and you have these Kannada and Telugu-based accents.
• 29:06
  SPEAKER_02: Hindi alone, you have what 300 million people speak Hindi in India, right?
• 29:10
  SPEAKER_02: Dialects and accents change from you know one city to another city.
• 29:15
  SPEAKER_02: Um that's like almost population of the US.
• 29:19
  SPEAKER_02: It's it's massive.
• 29:20
  SPEAKER_00: I mean, they I don't know that people really think about that per se because you're you're bought on it's almost like to a very minute degree of what you know the US looks at when you get closer to Canada or you get closer to the South or you get closer to California.
• 29:35
  SPEAKER_00: Like I remember traveling and I grew up in California, and people would would say, Oh, you got that surfer dude style accent.
• 29:44
  SPEAKER_00: I'm like, I don't know what you're talking about, you know?
• 29:47
  SPEAKER_00: And you almost get self-conscious about it to a degree.
• 29:50
  SPEAKER_00: And I think that's part of this as well is you know, depending on where you're at, it can actually, you know, detract from, like to your buddy.
• 30:00
  SPEAKER_00: I think bringing it back to your buddy's perspective, it probably made him super self-conscious about.
• 30:05
  SPEAKER_00: I mean, think about it.
• 30:05
  SPEAKER_00: He's he's at Stanford, he's he's um achieving and accelerating like a rocket ship in you know, in the first world of what you know, tech is, and then put him back into a scenario now he's getting you know profiled and slandered against because of his thicker accent, because of you know, culture that's you know, depending on who he's dealing with, right?
• 30:28
  SPEAKER_00: And it's sad, it's super sad, but you know, at the same time, you know, probably made him self-conscious about it.
• 30:35
  SPEAKER_02: Yeah, Chris, you're totally right.
• 30:37
  SPEAKER_02: And the sad reality, the sad consequence that comes from there is um whenever people feel or like any organism feels a mismatch from its environment, it adapts.
• 30:49
  SPEAKER_02: It changes.
• 30:49
  SPEAKER_02: That's just evolution, that's how it always is.
• 30:52
  SPEAKER_02: The same thing with people like Raul or people like me.
• 30:55
  SPEAKER_02: The way I sound today is very different from the way I sounded when I first came to Russia.
• 30:59
  SPEAKER_02: Same thing with Raul when he uh came from Russia.
• 31:02
  SPEAKER_02: When Raul came to the US, the way he sounds today is very different from the way he sounded when he first uh uh came.
• 31:09
  SPEAKER_02: So um, what happens is people change their accent to fit in.
• 31:13
  SPEAKER_02: And sometimes it's difficult for people who haven't experienced this bias to empathize with individuals that experience this.
• 31:20
  SPEAKER_02: Like I've experienced this, my family's experienced this, it's a very normal thing.
• 31:25
  SPEAKER_02: And so you can't really blame uh these individuals either, right?
• 31:29
  SPEAKER_02: Because really people are trained to understand the distribution of language or phonemes that they hear um consistently, right?
• 31:38
  SPEAKER_02: If you hear a new distribution of phonemes and you're not used to recognizing this distribution of phoneme, phonemes, um, which is an accent, um then uh you know your understanding starts reducing.
• 31:52
  SPEAKER_04: It's interesting, you sometimes uh watching a movie that would be set in you know Scotland or you know someplace in Ireland where you're trying to get used to the the accent and it takes time to kind of process that and and just begin to, you know, I always use the closed captioning uh in situations like that, but uh but it is fascinating because you you you know just getting used to listening to that over a period of time, you start to process a little bit.
• 32:21
  SPEAKER_02: It's a great example, yeah.
• 32:22
  SPEAKER_02: You know, people people turn on um uh people turn on uh uh captions for for movies, right?
• 32:30
  SPEAKER_02: With with you know foreign films or yep.
• 32:33
  SPEAKER_02: So it's it's a big one.
• 32:34
  SPEAKER_02: Um very good.
• 32:35
  SPEAKER_00: Especially, I think it's it's more powerful though, if you can um, you know, like listen to it in the native tongue, in the mother tongue, and then get the subtitles, which I get it.
• 32:44
  SPEAKER_00: You're all you're also creating a bit of bias because it's not how things are conjugated, it's not how things are are true in terms of what's being spoken and like all of the variations of let's say a single word.
• 32:58
  SPEAKER_00: Like I was reading an example, like how many, how many versions of um Russian words are there to describe the color blue?
• 33:06
  SPEAKER_00: Right.
• 33:06
  SPEAKER_00: And it depends on because like in in US, we don't have it as thick to have masculine and feminine as well, right?
• 33:14
  SPEAKER_00: So you have those aspects, and then I mean it's it's it's such a fascinating, you know, way because it hits it such a core area of just human connection.
• 33:24
  SPEAKER_02: Yeah, and it's interesting that we mention movies like these like massive streaming.
• 33:30
  SPEAKER_02: I mean, now that it's like a very populated market, but these uh streaming companies like they spend hundreds of millions of dollars on movie dubbing, right?
• 33:40
  SPEAKER_02: To make sure that I don't know, Brad Pitt's mouth uh moves the exact same way when or moves in the right way selling to you know move uh movies in China, right?
• 33:51
  SPEAKER_02: Or uh you know, movie is in China and alternates these other devices.
• 33:55
  SPEAKER_02: So um that's something our technology could also work on.
• 33:59
  SPEAKER_02: Imagine Brad Pitt's original uh uh voice but in another language with the right, it's a very interesting scenario.
• 34:07
  SPEAKER_00: I mean it's it's almost like a deep fake for voice, right?
• 34:10
  SPEAKER_00: Essentially.
• 34:11
  SPEAKER_02: And by the way, speaking of deep fakes, like whenever companies, um generative companies appear, um, you know, there's always like this other branch of companies that end up also formulating that help detect um, you know, whether this is happening, because um there will be a massive market for that too.
• 34:29
  SPEAKER_00: Yeah, it's wild west right now.
• 34:31
  SPEAKER_00: It's untap, you know, un you know, uh I don't know.
• 34:34
  SPEAKER_00: We we we've talked about it a little bit, but it's uh it's like I said, just crazy.
• 34:40
  SPEAKER_04: I just did a uh webinar with uh a school, a uh charter school, to talk to them about how to use generative uh AI and their data they work and how to how to deal with it in the classroom, which is fascinating.
• 34:54
  SPEAKER_04: What do you see as the downsides to this type of technology?
• 34:58
  SPEAKER_04: You know, people use things for nefarious purposes.
• 35:02
  SPEAKER_04: Would could you see someone using it to catfish or to, you know, commit fraud or or something like that?
• 35:08
  SPEAKER_04: And do you have safeguards uh around that?
• 35:11
  SPEAKER_00: Voice ID, right?
• 35:12
  SPEAKER_02: Yeah, before getting into that, are you teaching kids how to cheat in in school via Chat GPT?
• 35:18
  SPEAKER_04: Is that what's happening?
• 35:20
  SPEAKER_04: By teaching the teachers how to cheating and to teach the critical thinking skills about how they should be asking questions and also giving them DPT zero, which is the detector, right?
• 35:34
  SPEAKER_04: Around that.
• 35:34
  SPEAKER_04: So yeah.
• 35:35
  SPEAKER_04: So you know they're gonna use it, but you need to say what's the best way to use it, you know, and uh it can augment certain things.
• 35:44
  unknown: But yeah.
• 35:47
  SPEAKER_02: What's that?
• 35:48
  SPEAKER_02: If I had Chat GPT in college or even in high school or middle school, oh my goodness.
• 35:52
  SPEAKER_02: Yeah, very different person today.
• 35:54
  SPEAKER_02: So kudos.
• 35:57
  SPEAKER_02: Um but yeah, uh we have a very strict KYB uh process.
• 36:02
  SPEAKER_02: Uh each time we onboard a uh custom, you know, our investors emphasize that.
• 36:07
  SPEAKER_02: Um our customers emphasize that.
• 36:10
  SPEAKER_02: So very in InfoSec policies, right?
• 36:12
  SPEAKER_02: We have a bunch of these like three-letter acronym InfoSec um certifications.
• 36:16
  SPEAKER_02: So um it's it's really critical for us to be able to make sure that whoever we deploy with and whoever we work with are legitimate organizations and not you know have no nefarious um uh tasks or or you know actions.
• 36:32
  SPEAKER_02: Um so that's just part of our onboarding period, um, part of our onboarding process.
• 36:37
  SPEAKER_02: Um we um analyze our customers.
• 36:40
  SPEAKER_02: But you know, there are a lot of these companies that do this on an automated basis.
• 36:45
  SPEAKER_02: We don't work with them quite yet.
• 36:46
  SPEAKER_02: Um, I know like the mid-desks of the world and like these KYB companies.
• 36:51
  SPEAKER_02: Um, but I'm sure at some point we'll figure something out.
• 36:56
  SPEAKER_04: So where do you want to go with the tools and the technology?
• 36:59
  SPEAKER_04: What's on the horizon for you in terms of planning and applications?
• 37:05
  SPEAKER_02: For me, the importance of this technology uh really lies at its core um in allowing normal people to uh be able to um communicate, right?
• 37:16
  SPEAKER_02: There are obviously some limitations when it comes to identifying the synthesis to make sure it's not being used for nefarious purposes and all sorts of these edge cases, but um long term we think that uh the you know research world will and and the actual you know industry world will account for that.
• 37:34
  SPEAKER_02: Uh we want to bring this to enterprises, we want to bring this to um, you know, daily communications between people.
• 37:41
  SPEAKER_02: Um like we had a conversation with one of the founders of Twitch, and uh they were telling us that um, you know, they have a lot of gamers who are exceptionally successful in Asia and have trouble breaking into the US market because of their accent, the way they sound.
• 37:56
  SPEAKER_02: Um we want to build a system where people have a choice to sound however they want to sound, and uh we want to eliminate or not eliminate, or at least do something to improve um the bias that a lot of people experience on a daily basis and allow people not to feel the need to change the way they speak to fit in.
• 38:14
  SPEAKER_02: Really?
• 38:15
  SPEAKER_02: Like permanently, I mean.
• 38:16
  SPEAKER_00: I think part of that too is um if you guys understand the raw, you know, raw language coming in, then you can also help pick things up and also filter things out too, which could be good and bad, right?
• 38:32
  SPEAKER_00: But like I've I for instance, was working with a huge gaming platform company and we were looking for, you know, gamers that were grooming and that were um, you know, that were basically stalkers on kids, right?
• 38:45
  SPEAKER_00: Which is super disgusting.
• 38:46
  SPEAKER_00: But it was, you know, I felt it was so, you know, just that what we're working on, be able to catch, capture, and filter out, you know, these types of things so that you can, you know, uh make a safer environment, right?
• 39:01
  SPEAKER_00: Really, that's that's part of this.
• 39:04
  SPEAKER_02: I absolutely.
• 39:06
  SPEAKER_02: But this is one of those things that I kind of struggle with personally, right?
• 39:10
  SPEAKER_02: Like um, even though there are a lot of nefarious actors online, um, I also love America for its freedom of speech, um ability to express yourself.
• 39:22
  SPEAKER_02: And um whenever moving into that direction of hey, let's you know, censor, um, you always have to be careful about going too far into that direction.
• 39:32
  SPEAKER_02: And where do you really draw the line?
• 39:34
  SPEAKER_02: And if definitions change, right?
• 39:36
  SPEAKER_02: Like there are two ways to change the law, right?
• 39:38
  SPEAKER_02: You physically change the law or you change the definitions, right?
• 39:41
  SPEAKER_02: And if definitions change, the line shifts out, and you don't really know where where it ends.
• 39:48
  SPEAKER_02: So um it's an interesting scenario, um, but uh, I don't see Sonos expanding into that market quite yet.
• 39:54
  SPEAKER_03: Yeah, yeah.
• 39:55
  SPEAKER_04: I mean, so did you uh did you have anything that you can quickly uh uh demo for us or here?
• 40:01
  SPEAKER_02: So let me kick it off uh without Sonos.
• 40:05
  SPEAKER_02: And by the way I will intermediately enable and disable Sonos so you could hear how it changes.
• 40:11
  SPEAKER_05: Hi, this is Alex from the Customer Service Aid.
• 40:14
  SPEAKER_05: How are you today?
• 40:16
  SPEAKER_05: Great to hear.
• 40:17
  SPEAKER_05: I'm doing very well.
• 40:19
  SPEAKER_05: Thank you very much for asking.
• 40:20
  SPEAKER_05: So, how can I help you today?
• 40:22
  SPEAKER_05: Yeah, you can hear it.
• 40:24
  SPEAKER_05: I'm so sorry about that.
• 40:26
  SPEAKER_05: I'll be glad to help you.
• 40:27
  SPEAKER_05: Can I get your full name, phone number, and address to check in on that order?
• 40:33
  SPEAKER_05: Do you happen to have the order number so I can bring up your order?
• 40:37
  SPEAKER_05: Yeah.
• 40:38
  SPEAKER_05: Thank you.
• 40:39
  SPEAKER_05: Let me check on the status of your order.
• 40:41
  SPEAKER_05: Please give me a minute to check on that.
• 40:43
  SPEAKER_05: Thank you for reading.
• 40:45
  SPEAKER_05: It looks like your order shipped to the wrong address.
• 40:48
  SPEAKER_05: I'm so sorry about the mix-up.
• 40:50
  SPEAKER_05: Let me see what we can do about correcting your order.
• 40:53
  SPEAKER_02: You guys get the idea.
• 40:54
  SPEAKER_02: It's really cool.
• 40:56
  SPEAKER_02: And the cool thing is like literally on the application, you could toggle it off and on the exact same way.
• 41:02
  SPEAKER_02: Yeah, very cool.
• 41:04
  SPEAKER_04: It's very, very impressive.
• 41:06
  SPEAKER_02: Thank you for that uh impromptu demo.
• 41:08
  SPEAKER_02: So that was great.
• 41:10
  SPEAKER_02: Sorry, I couldn't make it live.
• 41:12
  SPEAKER_02: Um we'll have one of our reps call you guys and you guys could have a proper chat.
• 41:17
  SPEAKER_04: Yeah, they convert my accent into an Indian accent.
• 41:21
  SPEAKER_04: We're gonna take your thick Boston accent, Seth, and and uh yeah, make you more uh Well give me a Boston accent because I'm getting it.
• 41:31
  SPEAKER_04: I live in Boston, I'm from New York, but I sound like neither.
• 41:34
  SPEAKER_02: So we've actually done some like experiments with um uh accents within India.
• 41:42
  SPEAKER_02: It's also pretty cool, where we had uh uh one Indian accent um uh transfer to another Indian accent.
• 41:49
  SPEAKER_02: And there's actually just a lot of bias also within India, where um one person speaks with one accent and it's identifiable, uh one accent in English and it's identifiable by another person.
• 42:01
  SPEAKER_02: Um you could also modify accents there too.
• 42:04
  SPEAKER_02: And we've done experiments with that, and there's definitely a market for that too.
• 42:09
  SPEAKER_04: And it starts to, you know, at some point, um it is a translation of some sort, right?
• 42:16
  SPEAKER_04: You're translating accents.
• 42:18
  SPEAKER_04: Are you also looking at uh kind of the language translation uh as an extension or as part of this?
• 42:28
  SPEAKER_02: You know, not quite yet.
• 42:29
  SPEAKER_02: I think speech translation is a great space.
• 42:32
  SPEAKER_02: Um you know, and it's it's very interesting and very exciting.
• 42:36
  SPEAKER_02: Um but there are certain limitations that appear with uh durations, right?
• 42:41
  SPEAKER_02: For example, some languages put nouns and verbs at the beginning of sentences while others put at the end of the sentence.
• 42:48
  SPEAKER_02: Or you know, it it goes into some of these temporal based questions.
• 42:52
  SPEAKER_02: Right.
• 42:52
  SPEAKER_02: So if if a company is building inside of speech translation, um speech-to-speech translation, they really have to account for the permutations and combinations of different languages based off of grammatical structure and see how most commonly they structure things uh to be able to build a proper uh translator.
• 43:14
  SPEAKER_02: Um I think there are some markets like now where that's definitely the tech is there um 100% and um you um can really build off.
• 43:23
  SPEAKER_02: In fact, I've engrossed in some of these companies and um they're they're incredible.
• 43:28
  SPEAKER_00: So what do we uh you know, I'd love to get a take to in terms of uh Max, you know, what's what should we be looking for from you personally, and then what should we be looking for from Isana?
• 43:38
  SPEAKER_02: Yeah, you you know, we're continuing growing, um uh continually expanding our clientele, uh, focused on deployments and revenue.
• 43:47
  SPEAKER_02: Uh we're scaling out revenue.
• 43:48
  SPEAKER_02: So this is going to be an incredible year for us.
• 43:50
  SPEAKER_04: How many folks are you these days?
• 43:53
  SPEAKER_02: We're a 70-person team.
• 43:54
  SPEAKER_02: I want to expand out our horizons from call centers to more enterprises and building relationships there.
• 44:01
  SPEAKER_02: So um there are any enterprises listening in, you guys are always looking to reach out to me personally, uh, reach out to uh Sonus, and uh, we'd love to work with you.
• 44:11
  SPEAKER_02: Um our technology has a lot of different applications.
• 44:14
  SPEAKER_02: So even if you think of an interesting application that we may not have considered, you know, also reach out to us.
• 44:20
  SPEAKER_02: So uh um yeah, we're always welcome feedback as well.
• 44:23
  SPEAKER_02: So if you have any feedback on our demos, we'll also appreciate it.
• 44:26
  SPEAKER_00: I think I think this is a first for us too, though, Seth, right?
• 44:28
  SPEAKER_00: I mean, this is the first like you know impromptu demo that we've had on the podcast.
• 44:32
  SPEAKER_04: So we have not done that.
• 44:34
  SPEAKER_00: So Barty, set yourself aside.
• 44:36
  SPEAKER_00: Yeah, it's in this podcast.
• 44:39
  SPEAKER_04: Well, it's been wonderful to meet you, March.
• 44:42
  SPEAKER_04: Where are you based?
• 44:43
  SPEAKER_04: Are you in the uh Bay Area?
• 44:45
  SPEAKER_02: Uh I fly between uh the Bay and uh Bangalore.
• 44:49
  SPEAKER_02: So um you can find me one of those two.
• 44:52
  SPEAKER_04: Well, this has been fascinating.
• 44:53
  SPEAKER_04: I've really enjoyed it, and uh uh great to have you.
• 44:57
  SPEAKER_04: So uh thank you again for your time.
• 44:59
  SPEAKER_04: Chris, thanks for the introduction and um appreciate your your insights, and it's a fascinating area.
• 45:04
  SPEAKER_04: We definitely want to stay connected and hear how things evolve.
• 45:09
  SPEAKER_02: Seth, Chris, appreciate it, guys.
• 45:11
  SPEAKER_02: Let me know how it could be helpful and I'm here.
• 45:13
  SPEAKER_00: Thanks, my friend.
• 45:14
  SPEAKER_00: Appreciate the time.
• 45:17
  SPEAKER_01: Thank you for joining us on another deep dive into AI innovation.
• 45:21
  SPEAKER_01: Join us next time when we introduce another expert and discuss how to maximize AI in your world.
• 45:27
  SPEAKER_01: Early AI Podcast is sponsored by Early Information Science, CMS Wire, and Marketing AI Institute.
• 45:33
  SPEAKER_01: To learn more about us, visit Early.com at a Z A R L A Y.com.
• 45:40
  SPEAKER_01: See you next time.