The Five Nine: Can AI give telecom a new voice?

Mavenir adds Sanas Speech AI to its MAVcore voice AI portfolio

Deepgram acquires Pipecat creator Daily, debuts Real-Time Context Standard

Diana Goovaerts, Fierce Network: Ask anyone across the telecom industry and they'll likely tell you voice is dead. But is it really? A slew of vendors and billions in private capital are betting that AI can breathe new life into an old medium. Will these Dr. Frankensteins create an ungovernable monster or a new category of service operators can actually monetize?

Dag Peak, Alianza: So if you're an operator who can now very easily bring in an AI-powered application into your traditional telecom services and sell that high-value service to your customers at an uplift, right, then there's a real opportunity there to generate revenue monetize the network, and, you know, eliminate this criticism that voice is stale, right? 

Diana Goovaerts: I'm Diana Goovaerts, and this is The Five Nine. Today, we're looking at voice AI with help from Mavenir's Brandon Larson, Alianza's Dag Peak and Deepgram's Kris Efland to get three perspectives on what happens when AI stops sitting beside the conversation and starts participating in it. And we should really start by clearing something up. Voice AI is a broad label.

It can mean an AI agent answering a call, it can mean AI translating between two people while they talk or it can mean turning a finished conversation into information a business can actually use. Peak and Larson break it into three buckets: AI-handled calls, applications engaging in live conversations and post-call intelligence.

So this isn't just about replacing a person on the end of a customer service line. It's about what you can do before, during and after a conversation.

Brandon Larson, Mavenir: Voice AI is about adding AI into the voice path itself. So just to give you an example, some of the services we see as translation. We can have AI translate between two people that speak different languages. Basic use case, we've seen that launched. We can have AI be receptionists. So if you have a small business, and I call it the two jobs at once problem, think of a hair salon owner, right?

They've gotta service their customer while taking inbound calls from a customer, so we can have an AI reception answer the call on their behalf. We've seen about AI getting into calls to help with fraud detection. So, you know, the AI can be monitoring for deep fake voices, so voices coming from non-humans, as well it can detect scenarios like social engineering.

They detect urgency. You know, someone saying they're important, they're trying to get somebody to do something they shouldn't do expose information. So when we talk about AI voice, we're really talking about how do we take existing voice communications today and provide new use cases which provide value to the consumers and businesses using AI. 

Dag Peak: The fact of the matter is that human conversation is the richest way that humans communicate. This is how we bring minds together. It's how we form relationships. It's how we form trust. The fact of the matter is that we now have the capability of bringing conversations into the application and AI space.

And then once it's there, then of course, we get to take advantage of the astounding innovation and the astounding investment that's happened on the application and AI side of the equation. So it's the combination of those two things, the ability to get calls out to applications and then be able to do something with that call there's so many interesting things that we can do with voice, and that's what's resulting in this thing that we call the voice renaissance. 

Diana Goovaerts: For operators, that raises a bigger question. Is this a way to protect the voice business they already have or a way to build a new one? Larson's answer is firmly on the growth side.

He says product teams he talks to see an opportunity to add services without waiting for the next handset feature 

Brandon Larson: When we're talking to the product people inside the operators, you know, the B2C and B2B folks, they see this huge opportunity, said, "Hey, there's a lot of value we can create for our customer base by bringing these AI solutions in, and this is the first time we feel we can create value that's not completely dependent on the handset," right? What Apple and Samsung will or won't do. And when they're creating that value, then they can get uplifts, right? Uplifts and step-up plans or add-on plans. So they are definitely looking at this a way is to add to their top line by adding AI into the voice 

Diana Goovaerts: Peak, meanwhile, sees both offense and defense. In his view, operators risk losing customers to providers that can attach more intelligence to communications. 

Dag Peak: Operators, even if they're not aware of it, are losing subscribers, losing lines, losing SIP trunks to other vendors, like let's say Twilio, who can bring intelligence to that voice or messaging service. So if an operator can bring those services into their network, bring AI-powered services into their network, that's going to give them a better opportunity to hang onto their customers than they had before.

So that's the, the defensive play. But then that shifts right over into an offensive play and a revenue-generating play. 

Diana Goovaerts: Bringing these capabilities into the operator's own network and offerings could help keep the customer and create something new to sell. But something new still needs a business case, and Larsson argued that operators should start with a customer problem, estimate how many people have it and find out what they'd pay to solve it, then work backward into the technology.

Peak and Alianza, however, are taking a slightly different approach. They are building the plumbing that connects operators to a new kind of app marketplace, one focused on voice AI. 

And of course, none of this works if the conversation feels like a bad video call, awkward pauses, people talking over each other, an agent answering a question you asked two sentences ago.

Deepgram's Efland said the ingredients for successful voice AI include low latency, accurate speech recognition and reliable turn-talking. And accuracy gets a lot less forgiving when the conversation involves specialized language. For instance, a healthcare application, it's important to have AI be able to recognize drug names.

Kris Efland, Deepgram: If someone's telling you a specific drug name or whether it's the generic or whether it's the compound, you really have to know what the dosage is for that. You have to understand the pronunciation. You gotta understand like this generic actually means this compound name. So things like that really matter in like a safety situation.

Diana Goovaerts: But he also makes a distinction between getting a voice agent running and actually making it useful inside a business.

Kris Efland: The highest value that enterprises are gonna see is really getting that AI tightly integrated with their data and their systems. So we make sure that we are able to deploy wherever our customers are, whether they're on air-gapped on a device like a Qualcomm handset, or maybe they have their own infrastructure, their own VPC, their own cloud, maybe they have their own data center. We have to be really tightly integrated with customers. We spend a lot of time getting to know them up front, understanding what their use cases are, what their outcomes are, and then we work back from those outcomes. 

Diana Goovaerts: That is integrated with their data, their systems, the things the agent actually needs to know, and the tools it needs to do something about them. Otherwise, you've given the customer a very natural-sounding way to reach another dead end. 

Larson and Efland said connecting AI to the network isn't necessarily the hardest part either. Getting the model to behave the way you want in the conditions of a real call is actually tougher. Efland's example is a drive-through, right? A parent trying to order while kids in the backseat shout about chicken nuggets. That is not a quiet demo. That is Tuesday.

Kris Efland: We can cover the full range of human emotion, but I think it's making the conversations themselves work reliably in unpredictable real-world environments. That means, you know, I'm pulling into the drive-through with my four-year-old in the back or five-year-old in the back shouting 'cause she has her friends from her soccer team. They all want chicken nuggets and I'm trying to order and the drive-through just gets it perfect, right? I think those are challenging conditions that our customers are facing.

So in that scenario, the context is, "Oh, I'm a drive-through. Oh, there's multiple speakers. Oh, I know that there's probably one that's closest to the speaker. Oh, I know there's probably someone who's sounds a little older maybe than some folks that sound maybe a little on the younger side. What does that mean? I don't know, right?"

It should mean something. And so this context starts to roll starts to accrue and starts to add a ton of value in terms of how these a- how this AI is thinking about the problem.

Diana Goovaerts: Larson also singled out context engineering as an under-appreciated challenge. His argument is that knowing a subscriber's preferences and history can matter more than simply choosing a more powerful model. Peak is a little bit more bullish on how much remains unsolved, though, and he said the major problems have been identified and sees the pieces coming together.

But even in that optimistic view, trust, privacy and consent remain central to getting people comfortable with these types of services. So our guests didn't quite agree on how much of the work is finished, but they do keep coming back to the same test. Does the whole system work for the person using it?

And that is where we start to get to the part that can't just be hidden behind some smooth voice: permission. If I turn on an AI service, I'm giving my consent. But what about the person who's calling me? What have they agreed to? Do they know whether the conversation is being processed, recorded and used later?

Larson describes a service design where the other caller hears an announcement and actively accepts before the AI starts. Peak said Alianza's trust framework handles consent, announcements and collection alongside vetting the experiences attached to calls. Those are their descriptions of their approaches, not a universal legal rule for every service or every location, but the shared design challenge is clear: Explain what's happening without making the caller sit through an obstacle course.

Kris Efland: Generally speaking, you know, users need to understand if there's AI participating in a conversation whether it's being recorded. I think you're hearing a lot more of the disclaimers happening where AI is listening and, you know, they're getting better about consolidating the, the legalese. But anytime the data is processed or transcribed, I think we owe it to the consumer to be transparent about that.

Diana Goovaerts: There's another layer here too, though. Agreeing to use AI on a call doesn't answer every question about what happens to the data afterward. Larson describes the operator as a steward of subscriber information, while Peak argued the participants are the ultimate owners and said privacy notices need to explain how the information will be managed. 

Dag Peak: Once that consent has been granted what you must do is tell the participants what they've actually signed up for, right? Which is all about how you deliver a privacy notice that tells the participants of that conversation what they actually consented to. 

In those privacy notices, this is where it will be defined who is actually responsible for the content, who can manage the content, who can see it, who can't. All of the nuances of how that content can be managed over time get defined there.

Diana Goovaerts: And Efland said that in some cases, Deepgram uses customer data to improve models where enterprises want those improvements. He also said customers can opt out and that some enterprise arrangements provide for zero data retention. The takeaway for anyone buying these services, ask separately about participation, recording, retention and training. A useful conversation and a reusable training data set are not the same thing. 

So where does all of this stuff actually run? Larson makes the case for bringing AI into the operator's network, citing control over the experience, latency and costs. 

Brandon Larson: Bolted on the over-the-top app, and by that I mean if we just kinda have connecting the network to, say, cloud technologies like OpenAI, right, for example. You're gonna deal with the connectivity, the latency, and then you're dependent on OpenAI's charging model for how you charge in the network.

So bringing AI into the network means that operators don't have to worry about the privacy, right? It's behind their firewall. They own it. They own the user experience. They can manage latency, and most importantly, they control the cost model, right? 

Diana Goovaerts: Peak's answer is more dependent on the application. For real-time translation between two people, he says putting the application close to the call's media streams has major advantages. But for an agent answering the phone, he said outside applications can still be responsive enough. And if you're summarizing a call after it ends, well, you have much more room to work 

Dag Peak: There's a spectrum of how performance sensitive the various AI voice use cases will be, and that will really drive where you ultimately wind up deploying them.

Diana Goovaerts: In other words, put AI into the network isn't a whole strategy. You still have to ask which AI is doing what job with what performance requirements. 

And then there's the thorny issue of pricing these new AI services.

Brandon Larson: I do think they're, they're looking to say, "If we stay with that flat pricing model, right, can we upcharge another $5 a month? Or can I get you to a higher tier plan?" So instead of $40 a month, you'll pick the $60 a month plan, which they call step-up, because it's got more services and features with it.

That's one model of thinking, and that goes with the current paradigm, which is we're just gonna try to get people to you get into higher plans. 

The usage-based model may be tricky because you get the taxi meter effect. So there is a detractor to go back to the usage-based models because then people, as a consumer, it's like, "well, how much am I spending on this call or on this interaction?" And then that starts detracting from the user experience. So I think that's what they're trying to weigh. 

Diana Goovaerts: So there is a bit of a balancing act happening. You need to build a service that feels easy to use with economics that still work when people actually use it. Peak says a win for voice AI is a customer who sees the operator as someone solving a business problem and not just selling a connection.

Dag Peak: Joe the Plumber does not care one single bit if that call came to him as 4G or 5G or 6G or 27G, he doesn't care. He wants that call to be handled. And so a win is Joe the Plumber thinking about my service provider is solving those kinds of challenges for me, right? They're not giving me just another iPhone. They're not giving me just another upgrade in my data plan, but they're solving real business challenges for me in a way that wants me to stay with them, in a way that wants me to pay them for helping me solve these challenges. 

Diana Goovaerts: Larsson's measure of voice AI's success is revenue growth alongside customers seeing enough value to stick around. Efland pointed to new services, new revenue and measurable business value. For the person making the call, his vision is simple. Get the problem resolved without being transferred from department to department.

And voice AI actually doesn't have to stop at the phone system. Asked for an example beyond customer service, Efland imagined a museum where you can talk to the artwork, ask about the artist, the story behind a sculpture or painting, or hear all of this information in your own language. 

That's a different kind of voice service, and it's a reminder that the opportunity here is bigger than putting a friendlier voice on the same old menu.

Kris Efland: I think one thing that I hadn't even seen done in science fiction before that I thought was super interesting was what if the artwork could talk back to you, right? So you're looking, you're at a museum, there's a beautiful sculpture. Maybe I wanna hear the story about the battle that it's telling about. Maybe I wanna hear about the artist from the artist themself. What did they say about this piece? Like what was it commissioned for? 

What used to be a painting on the wall is now like, you know, still somewhat like physical but very flat and very two-dimensional. Now you're giving it a ton of depth through a whole new experience that can be either personal or it can be shared, right? 

Diana Goovaerts: For telecom operators, though, the immediate challenge is very grounded. Pick a problem that customers actually have. Build or have ecosystem partners build a service that can handle the messy conversation and not just the scripted one. And find a price that works on both sides of the call. 

Because a voice that sounds human might get our attention, but a service that actually understands what we need and reliably helps us get it done is what could actually make us pay.

That is all the time we have for today. Make sure you like and subscribe wherever you are listening or watching, and stay tuned to fierce-network.com for news and insights during the week in between episodes. Catch you next time.