All Podcasts

VOTF Episode 4: Mining for Meaning: How AI is Extracting Conversational Gold

Host and Guests

Host

Sam Zegas

Sam Zegas is life-long language aficionado, with years of study in linguistics and foreign languages, and now he can add “podcast host” to his resume. He holds an MBA from Harvard Business School and an MPP from the Harvard Kennedy School of Government.

Guest
Todd Fisher, CEO, CallTrackingMetrics

Todd Fisher, CEO, CallTrackingMetrics

Todd Fisher is co-founder and CEO of CallTrackingMetrics, a global conversation analytics platform that enables marketers to make data-driven decisions and increase ROI. Todd founded the business in 2011 with his wife, Laure, in their basement and together they have grown it into a Inc. 500-rated, award-winning solution serving over 100,000 customers around the world.

We’ve definitely realized that the robots are not coming for all of our jobs and we have a long way to go before we successfully automate all things mundane. BUT – we are starting to see AI’s power when it comes to extracting insights from long form conversations, according to Todd Fisher, CEO of CallTrackingMetrics. Tune in to hear how AI isn’t just hearing us, but is listening and mining for meaning that can improve our lives.

Read Full Transcript

Sam: Welcome to Deepgram’s Voice of the Future podcast aka our favorite nerds. At Deepgram, we’re obsessed with voice, and this podcast is our exploration of the exciting emerging world of voice technology. I’m your host today, Sam Zegas, VP of Operations at Deepgram, and our guest is Todd Fisher, Co founder and CEO of Call Tracking Metrics. Welcome to the show, Todd.rnrnTodd: Thank you. So good to be here.rnrnSam: So we affectionately call our podcast our favorite nerds, and of course you’re one of them. So maybe you could start by telling us what kind of a nerd you are.rnrnTodd: That’s that’s awesome. Well, so I’m I’m a kind of nerd, I guess, that I really like to solve problems. I’d like to learn new things. sort of I’m driven to just learn more about, like, what’s going on in the world and, like, how, you know, things that we are doing in software can can apply, but also in the physical world. Can can can solve problems? kinda like to get my hands dirty. It’s it’s fun. And, yeah, I think that that kind of is the kind of nerd I am, I think. I like to like to get in there and solve a problem. So I work best when there’s a problem to solve.rnrnSam: Yeah. You were telling me earlier that you in the last few years have gone into learning about electronics and circuit boards and things like that. What does that journey been like for you?rnrnTodd: Oh, yeah. Yeah. It’s been it’s been pretty neat pretty amazing. So so, you know, I’ve been a software engineer pretty much my whole career. since the very early days of my life, been doing things on on computers and software. And I’ve always been, like, fascinated by, you know, electronics, but I never really took the plunge. And it was kinda neat also in my forties to be like, okay, I’m gonna learn something new. Right? because it’s like, you know, just been a software engineer forever. And so knowing how to write programs, just kind of, like, there’s, like, writing English or something. And so it was, like, really hard to, like, open up a data sheet and be, like, what is this chip? And how does it work? And what’s a capacitor and what’s a resistor? And how do they all interact with each other? And, you know, what’s difference between positive and negative? And And and just sort of, like, to once again, go back and and go through these hurdles of learning something new and having that experience of, like, like, not knowing anything and being like, oh my god. I’m gonna, like, how do I learn this? And so to me, that was, like, really kind of the coolest part, you know, of two years in and kind of on, you know, nights and weekends. Right? Night and weekends warrior? What is it a weekend warrior. So, you know, because with this job, there’s not that much free time. Right? So little bits of free time here and there, read the datasheet, watch a YouTube video, you know. YouTube University is what a lot of folks like to call it these days. I went there. It was very enlightening. And, yeah, now we have like, what we call we’ll we’ll be releasing soon CTM powered status lights which are fully integrated circuit boards that, you know, you can essentially attach to a desk and it will automatically without being connected to your computer. change the light color based on the status of the agent. And hopefully, it’s pretty much hands free. But one thing I’ll tell you is, like, anything. They’ll, you know, they’ll they’ll be glitches because it’s software. There’s things to learn about that, you know. So that’s really interesting when those two different parts of your life collide that way. And kudos to you for going back and trying to learn something brand new as Nicole. I’ve done that and it is it is really uncomfortable to have your comfort zone of knowledge and then suddenly to leave it, it’s a it’s it’s a challenge for sure. It’s crazy. It’s like, whoa. Like, what is this new stuff?rnrnSam: Yeah. Well, great. Why don’t you give us an introduction to Call Tracking Metrics? What you do? What your mission is? All of that.rnrnTodd: Cool. Yeah. So so call tracking metrics. Call call tracking metrics is sort of like one of my children. It’s u002du002d Yeah. u002du002d it’s it’s it’s it’s growing and it’s always evolving. And, you know, we we started it around twenty eleven, twenty twelve. And you know, our mission is is is really to provide communication software and and marketing attribution to all all types of this. Right? And what we have found is that it’s really best if we sort of approach it as a as a how can we help that business operate better? Where are the opportunities for efficiency with respect to communications? And and how do we help glean insights from that communications with the businesses customers to help them optimize, you know, their sales, their marketing, they’re targeting. You know, their automation around, you know, understanding whether or not this was a good or bad call, but you’re really getting into the details of what were some of the details that are meaningful about that phone call? And can we extract that out? And can we get that into your CRM in a in a efficient way that doesn’t require you to do a lot of manual, you know, typing and or bookkeeping. I find that, you know, sometimes, well, going going way back to when I started, somebody brought a spreadsheet to me and said, hey, this is what I do to keep track of my my child’s health care. This is back in two thousand. And I we said, oh, well, let’s turn this series of spreadsheets into software. And so I find that still, today, we’ll encounter a business that has been keeping track of their leads in a spreadsheet, and they’ve been, you know, following up by calling through a spreadsheet or keeping track of conversions and sales and especially, and, you know, any opportunity for us to kind of work with that business and think about how can we automate that and and and alleviate the burden so that they can focus more on the interesting conversations with their customers and and and, of course, help their customers even more, right, in the process.rnrnSam: Mhmm. So you’re you’re really a platform that provides sales enablement and tooling so that you can automate marketing attribution and increase the performance of sales teams. It sounds likernrnTodd: Yeah. Yeah. I mean, you know, it’s it’s interesting because I would say it is a journey of discovery of who we are and how we can help our customers best, and that is probably the best way to describe what we are today to anyone that comes to seeking us out. I think we’ve become pretty darn good at this through a number of features, but mostly it’s it’s all been driven by you know, conversations with customers around how can we help them. Right? Understanding their particular needs has always driven us and probably we always will drive us.rnrnSam: So That’s great. Of course, there’s a lot of language processing that goes into doing that sort of work well, both in the transcription side, in the speech understanding side, understanding the intent and the meaning of what people are talking about. Are you mostly doing real time use cases? Or are you doing things that are asynchronous as well?rnrnTodd: We’re we’re doing we’re doing both. Yeah. So we’ll be doing real time and and asynchronous. I’d say the vast majority of our customers probably rely on the asynchronous side of of the analysis. Simply because, you know, when you think about, like, processing an audio file to determine whether or not there’s a stalemate or not is not typically something that has to be immediate. So it’s often that it just makes sense to allow the call to finish, process all the things, and then you can also you can look at both the the the understanding of the what was said in the call and compared to the other facts that you may have collected. Right? So if there was an agent on the call, did they record something into a CRM? And and can we, you know, align these things together to get a better picture and then make a decision about how to send this up to, say, a Google Ads for conversion tracking. Right?rnrnSam: Yeah. That’s fascinating. One of the trends that we really watch at Deepgram is the idea that actioning from long conference stations will become more and more usable as time goes on as accuracy improves as our analysis of intent and and your analysis of intent gets more sophisticated, the idea that a machine could listen to a long conversation and then take action automatically as a result is just a really promising future.rnrnTodd: Yes. I think it’s it’s phenomenal. And I think that, you know, it’s I would reflect on the fact that, like, back in twenty nineteen, I’d say, you know, if you look at the sort of the the way things were going in twenty nineteen, a lot of folks were very, very concerned over, you know, oh my god, we’re gonna, you know, automate away all the jobs. Right? Mhmm. And now fast forward to, like, you know, now, and it’s like, we don’t have enough people to do all the jobs. I wish we had more automation. Right? And so I think that the more and more we focus on, you know, having meaningful interactions with other people. Right? Like, what we’re doing right now and and taking out and not having to worry as much about the the manual process of having somebody else listen to our conversation, like, for, you know, hours and hours on end, and instead allow computers to do the sort of the manual work of of listening and extracting the meaningful information, then I think people like us can and and people around the world. Right? Can can have more meaningful interaction and accomplish more. Right? So actually, it increases productivity, and it allows people to to do more, and and and that’s just a good thing in general. So I agree that that is an interesting connection there because I too have been hearing for a long time about how automation will reduce the need for human workers and get it actually just creating new and different opportunities for human workers to be more productive or more engaged. Yes. Yes. which I think that leads to a better quality of life too for everyone involved. Right? Because if you get to to use your brain in different ways, right, I think it’s a more interesting exercise at the mind and most likely keeps your day interesting and maybe goes faster. Right? Yeah. So it’s a good component.rnrnSam: Yeah. How would you say you’re different than your your competitors? What’s what’s different about you?rnrnTodd: Cool. Yeah. So I think I think it’s interesting because we compete with a lot of different types of businesses. Mhmm. But what’s probably most unique about us. I’d say, number one, at least I think this is unique about us. It’s probably our approach to our customer. But moreover, it’s what we provide as a software solution. And that is we we provide both the the communication tools integrated into one platform and the marketing attribution system. So when you think about a a typical call tracking company, they might be thinking, about just how do you get the attribution data from a phone call to an AdWords or Google Ads nowadays. And I think that While that’s a very interesting problem to solve, I’d say that it’s probably been solved pretty well over the last decade. And so what’s more interesting now is how do we take all the data and and provide you with a full experience. Right? End to end experience. So You know, you can go out, you can buy a very expensive phone system, which has all the Wizz Bang reporting features. You can then go out and buy the attribution platform, which has all the Wizz Bang you know, reporting capability, or you can just buy a call tracking metrics and it has it all included. And it’s all integrated together already with workflow automation, a wonderful integration with Deepgram that enables many, many types of voice and speech analysis and signal processing that I think is really, really top notch. Thanks to Deepgram’s amazing enough. speech recognition capabilities.rnrnSam: So it’s great to hear it.rnrnTodd: Yeah. I mean, you know, promote, but, you know, it’s it’s true. I I should share, like, you know, before working with Deepgram, you know, we’ve worked with well, first, I tried to roll my own because in in university, our studied machine learning and had a pretty good understanding of how we could do, you know, speech speech to to text processing, and so did a lot of a source work on this, ultimately decided, you know, this is gonna be a lot of manual labor, a lot of effort to build a really good machine learning model. we have a lot of things to focus on. And so it’s it sort of speaks to the value of what we provide, which is we pull together the best vendors and capabilities into one platform and offer them up in an automated way so you could, you know, you could program our platform basically. without being a programmer.rnrnSam: Yeah. That’s great. We are honored to be part of your stack. It is a it’s a very difficult technical challenge to take an unstructured data type like audio and turn it into something structured like text because there’s so much variability and and idiosyncrasy in the way that people generate speech audio that you then need to collapse down. It’s a it’s a very interesting challenge and one that we love to work with every day.rnrnTodd: Yeah. It’s really cool.rnrnSam: So I’m curious what? What is the hardest part of the work that you’re doing today? What frontier are you trying to push out these days?rnrnTodd: Oh, I think I think probably the hardest part right now is is scale and and not so much scale in terms of the customer base as much as it is scale in terms of the types of problems we’re trying to solve with respect to automation. So we want to make it easier and easier for our customers to never have to do anything manually. Right? I feel like or repetitiveness. Right? We wanna kind of eliminate all the repetition. So I’ve been telling our teams and and folks on the team, you know, anytime you encounter especially in our professional services team. Anytime you encounter something where you find that you need to use something more than one time, you know, bring it to my attention, let us know so that like, that can be our focus. Right? Is how do we eliminate multiple clicks to accomplish a goal, but more so, multiple same clicks to accomplish the same goal for many things. Right? Like so we we, you know, we have add agencies with us that maybe have thousands of accounts and thousands of small business customers or big big customers. And I want to seek out a way for them to never have to do the same thing more than once. Right? because that would be insane. Right? So we’re trying we’re trying to minimize that, and I think that’s one of the hardest things. And part of that is also we’re we’re trying to cater to more markets, more more international markets. And so getting into these different markets and working with people abroad is very interesting. Never never would have dreamed of of what that would have been like, but here we go.rnrnSam: So Yeah. So you have multiple languages in scope now?rnrnTodd: We do. Yeah. We do. Yeah. And so, again, it helps that Deepgram has the different trans the ability to do transcriptions in different languages, it’s really, really valuable. Yeah. Yeah. It’s a and that does a an interesting challenge in itself as we think about how we approach new languages, anything that you do for English related to analyzing intent or working on summarization or something like that, all of these different things that we research it’s it’s completely different. Every time you start doing it, do you really need to start thinking about how you label and tag and train to do those sorts of tasks in that language other than English?rnrnSam: Yeah. Imagine that’s a, you know, that’s a big hurdle.rnrnTodd: Yeah.rnrnSam: Yeah. So how does machine learning fit into your solution?rnrnTodd: In in general? Where does that come in? Yeah. It’s a it’s a great question. So so, generally, I I try to look at it in terms of what problems can be solved. Right? So there’s there’s a lot of problems we can solve by just looking at the data and and making decisions on the data. Right? So machine learning should give us the ability to kind of, like, look for maybe patterns across the multiple data points classification. So we think about, like, what was just said in this snippet of a conversation? And can we classify that in a sort of a a meaningful or actionable way? Like, you know, perhaps in this snippet of text, they indicated that, like like, for example, can can you answer the question of did they provide you with the appointment date? Right. So that would be, I think, a good use case for us to invest in machine learning. Right? Because how you might express the confirmation of an appointment date may be different than how I might do. Right? And so I think a machine learning model could be really good at that kind of problem. But then the subsequent follow-up problem is more like, is that appointment date a time in the future? And is it a time in the future that is reasonable? Right? That we can probably solve with that machine learning. Right? Because we can just say, is it greater than now? And is it less than maybe some threshold of realistic future appointment date? Right? That would be relevant to business that is collecting that appointment date. Yeah. That was okay. So that’s the kind of nerd I am. I like to think about all the, like, like, stupid details. Yes. Someone’s got a yeah. I know that it’s it’s really quite important. So you’re you’re using machine learning for conversational analysis type tasks.rnrnSam: Yes. You know, I I think anyone who is involved in machine learning these days should be thinking a lot about bias. Most people do have some sort of question going out about bias. How does bias come into your consideration of ML?rnrnTodd: I think that’s that’s really cool question. So so currently with the way I’m thinking about it is that we’re kind of thinking about it at a very simple level. So we’re thinking about it in terms of, like, small snippets of text. Right? And so I haven’t really thought that bias could inflict its eagles into those little So it’s well, primarily because the use cases that we’re we’re focused on right now would be intent based. So, like, you have expressed the appointment date to me. Right? Or you’ve expressed your account ID to me? Or you’ve you know, express for willingness to purchase. Right? I do think that one of the ways we might solve Some of the biases though might I’m thinking maybe the emotion detection capabilities of Deepgram could play a role in this. Right? Because if it has the ability to maybe detect sarcasm, which I’m not sure if that’s gonna ever happen, but if if we have like sarcasm kind of capabilities, then I I would think that that could play a pretty handy role in identifying you know, the sort of the the misinterpret of of of a of a of a of a snippet of text, right, per se. Right? If we knew that, you know, this was spoken in anger, right, versus spoken in sort of normal terms, then then maybe we could flag that in some way and say we’ll we’ll we’ll consider that in the in the in the yeah. At some point, right, you have to take whatever it was that was classified. and how is classified and and decide what what should we do with that new knowledge. Right? Should we respond with something or or not. Right? So I think maybe it becomes a variable that you could utilize to say, okay. Well, in that case, I should respond this way. Right? It’s such an interesting topic. And, you know, broadly under the umbrella of things that Deepgram is actively researching we we think a lot about different categories of speech understanding. Can we summarize? Can we identify a topic? Can we understand the sentiment? And some of these things are already available, by the way, if anyone goes to console dot deepgram dot com, you can check that out. But as you say, there’s there’s so much variability. Right? sarcasm coming from one person may sound really different than it does for another person. Or sarcasm in English might take on a different form than it does if you’re being sarcastic in Spanish. And these are things that we have to to study and build models that can perform in a way that’s actually producing useful results for people who engage with those models. we’re also always thinking of bias in terms of representation of whose speech data gets into our models, making sure that we can cover all sorts of meaningful demographic slices so that anybody could submit audio to Deepgram service and get a good result back. But it’s a very difficult challenge for sure.rnrnSam: Yeah. Absolutely. Yeah. Yeah. The the problem of bias is definitely yeah. That is that is especially for training models, I think that is that is where is is really infectious. Right? This is when you’re in that training phase. Absolutely. Yep. So I wanted to move on to another really interesting conversation topic that you and I had recently around closed loop attribution. I think it came up in the context of if someone doesn’t keep up with the latest in your space, how might they get left behind? And I’m curious if you could just expand on that topic for a bit.rnrnTodd: Sure. Yeah. So I I think I think on the on the on the closed loop attribution side and, you know, if they’re not, you know, sort of keeping up to date on on on on these platforms, you know, you’re really gonna be able to leverage software, like, you know, Deepgram and and Call Tracking Metrics in order to you know, end to end track and and provide feedback. You want you want this feedback loop. Right? You wanna be able to know that someone someone called your business or maybe filled out a web form in your business, then they called, maybe they maybe there’s a follow-up text conversation. But ultimately, the the the data points that led to the sale, you wanna know when that occurred and and what maybe click IDs, let’s say, from Google, brought that person to you. Right? And then the journey there may go through a CRM where maybe the sales finally closed. And then when it’s closed, you you wanna be able to select the click ID that’s most relevant to the attribution model that you have select for your business so that you can inform Google that this is the click that prop, you know, your your business in resulting in maybe you shuffling around your advertising budgets so that you get more high quality individuals like that. Right? Mhmm. So, you know, it sounds very simple, but the mechanics and the automation required to make it happen at scale is is I think where you know, a lot of attention can be given to business to improve improve the growth of the business. Yeah. So hopefully that makes sense. It does. Yeah. But there’s such a big ecosystem that you’re working in there connecting into ad tech and things like that. you you probably have I would imagine you have quite a big team working on how you engage with a system that’s not complex. Yes. Yeah. Actually, well, I’d stay in in the space called tracking metrics probably has one of the smaller teams, but maybe the more talented teams, I don’t know. So u002du002d Lovely. u002du002d that’s that’s it. Yeah. We we we run pretty lean and mean. just because I think it’s it’s a good way to do it and keeps us honest. Yeah. I can really do that.rnrnSam: We try to do the same at Deepgram to efficiency is is definitely one of our top priorities here. Great. I I wanted to ask if there’s any events or opportunities to connect with you or your team coming up this summer that you wanna make the listeners aware of.rnrnTodd: Oh, sure. I appreciate that. So we at Call Tracking Metrics are gonna have what we’re calling a very googly July. where we’re gonna be announcing many new features in with respect to how we integrate with Google, the ads, Google Analytics, Google Business Chat and some other capabilities in there along the way. And we’re pretty excited that yeah, are googly July, and we’d love to have as many folks as possible around to kind of participate. So yeah. think it should be pretty exciting.rnrnSam: Fantastic. Well, cool. We’re almost at the end here before we go. Okay. Let’s take a minute to remember just how far technology has come even in our lifetime. So we always close this way. I’m gonna ask you to explain a piece of outdated technology like he would to a kid who was born after twenty ten. So this person is ten to twelve years old, and they’ve lived their whole lives in the era of the iPhone. Of course, the size and the power of our devices is in part dependent on chip size. So I’m curious how you would explain the impact of chip size to someone born after twenty ten.rnrnTodd: Yeah. I mean, I would say, you know, that device you’re holding in your hands that you actually probably never removed from your hand and and or wrist. Right? Is only possible because the chip sizes have gotten so so dramatically small much smaller. The the the the the two eighty six desktop PC that weighed probably fifty pounds when I was that age is less of a computer than the iPhones they carry in their hands. today. Right? Pretty amazing to think about. And, yeah, I I mean, chip sizes it’s an incredible thing, how much smaller they become, what it means in terms of what we can do. you know, it’s it’s easy to think, you know, people have said for a while, Moore’s Law is over, but there’s so many ways in which it is not, that it’s just tremendously amazing for the future of the world and just what we can do. So I I think there’s a lot of things to be optimistic about with respect to how that space is evolving and evolving so so fast. And and it really does mean that, you know, the the the artificial intelligence kind of capabilities are are possible now because of that silicon becoming that much more efficient if it’s even silicon anymore.rnrnSam: Yeah. That’s great. I love to end on that note of optimism. Well, ten year olds around the world are thanking you for your explanation. And while the account is missing out there, it’s amazing just to think about how far technology has come even in our lifetimes. time. Thanks so much for being one of our favorite nerds. It’s been great talking to you today. Likewise. To all our listeners out there, thanks for tuning in. come back or come check us out from our info. We are obviously Deepgram and Call Tracking Metrics. That’s call tracking metrics dot com. And, of course, you can find us at deep deepgram dot com and at deepgram AI across all of our socials. So with that, we’re out. Catch you next time.