Conversations on Voice In a series of three articles, we’ll be sharing some of our thoughts on voice-first projects. To consider concretely what it means to build a voice-driven product today, we’re kicking this series off with wysp founder Mark Larratt.

January 14, 2026, By Bakken & Bæck and Mark Larratt

With the integration of new AI functionalities, the world of voice-driven products is rapidly opening up to the infinite applications of natural language processing (NLP). As a studio that has been building digital products for more than a decade, we’ve gathered first-hand expertise in the various ways this technology can be applied across industries.

In a series of three articles, we’ll be sharing some of our thoughts on voice-first projects and products. Trying to create a broad understanding of how the technology and field as a whole are evolving, we'll look at both from different perspectives, interviewing our colleagues working in product, development and branding about their current approach to voice projects.

To consider concretely what it means to build a voice-driven product today, we’re talking to Mark Larratt, who, in partnership with BB, has built wysp [link to wysp]—intuitive, voice-only navigation. We’ll consider some of the implications of using voice as a medium and how this translates into the design of a new product.

How did you arrive at the idea of developing a voice-first product? Have you been interested in the technology itself for a long time, or is it a means to an end?

It's always a means to an end. I do get distracted by shiny things, but the period in which I got distracted by voice just because it was shiny was over 10 years ago. Being able to do things with voice is pretty old news, but being able to do it really, really well feels like it's just becoming possible now. So, it's back on my radar in that way.

I came to this project during my experience delivering bagels around Oslo. It wasn’t a set route, like a postal service might have, but followed a similar pattern to most other delivery apps that accept orders as they pop up. I'd had enough difficulty with keeping on top of where I was supposed to be delivering, how to get there, where I was in terms of schedule-calculating, wondering if now would be the right time to stop and do this other chore, take a break, whatever it might be.

I found myself misusing dangerous amounts of attention while driving. Your eyes aren’t on the road all the time, which is maybe fair enough when you’re road-tripping or just doing some casual navigation, but that changes when you're navigating between random points all day, every day in dense urban areas. Even with a mounted screen, any interaction deeper than a quick glance still requires stopping the vehicle to be done safely.

So you wanted to address these needs, and found that navigation is something that lends itself well to voice technologies?

Navigation stands out to me as a rare example, in that it has experienced mass adoption of voice features for decades already. You generally can't navigate exclusively through voice, but it works well enough to be a standard interaction. Depending on the kind of navigation you’re doing, you end up having to disambiguate directions by looking at your map to find your GPS dot when it says to turn left, for example, or checking if it’s this street and not the next. Any change of plan requires diving into the interface to reconfigure the route.

My thought was, “Could I make voice navigation work so well that you could use it voice-only and not have to use a screen at all?” That means it has to be extremely clear, extremely intuitive and extremely timely, so that you don't have to be looking at any graphics to disambiguate it.

Once you've defined the concept, how do you start defining the actual voice element of the product?

For me personally, my basic assumption is that when I'm using a system like this, I don't want it to talk at all. My goal is not to hear it talking; my goal is to get to a certain place in the city, on time, without making a mistake. And when I arrive, I want to know where I'm going and who I'm going to talk to and what buzzer to push and what I should have in my bag for them. That's what I want.

Originally, routing for something like this would only consider travel distance— you take the shortest route to the place. Over the years, it’s become more sophisticated and takes traffic, tolls, and other preferences into account. Wysp works similarly, but it adds an additional value: it cares that the route is easy to follow and doesn't put you in tricky situations, traffic-wise, potentially at the cost of a slightly longer route when the schedule allows it. It prefers routes that are easy to successfully communicate without having to micromanage you with constant talking.

The communication is designed to feel comfortable even during long silences. During departures it will quickly brief the driver on the duration and general direction of the next leg, and perhaps one or two landmarks or major road choices, so that the driver can subconsciously feel that they would notice if something had gone wrong. If I feel like I know the general plan, a quiet system disappears from my attention. If I feel unsure about it, then a quiet system just feels broken.

From there it's about trying to predict the moment when someone is going to want to know something. If it speaks up and says something you're not yet curious to know, you might not even hear it – you perceive that something is being said, but it just kind of floats past you. But if you're primed, and starting to feel curious about what's next, that's the perfect time. This concept of salience is pretty crucial in voice interfaces. If the voice is too“attention-grabbing”, then it’s grating and annoying. If it doesn't grab enough attention, then it can be misheard or ignored. It has to feel good to hear from it, too, so that leads into the design of voices and the tone of voice.

We’ll be talking about branding and tone of voice in a later Q&A, but from a product perspective, are there other ways you can try to personalise the voice in this kind of setting?

The goal is to create a skilled communicator that feels confident and precise. I want it to feel like you're being guided by someone who has lived in the city their whole life.When a navigation system mangles pronunciation, that expert feeling falls apart. We've been able to implement a system of code-switching, so it will speak to you in whatever language your phone is set to but still pronounce all local nouns correctly(ish). I'm also seeing good results from using voices that maintain a consistent local accent, while still keeping good pronunciation in mixed language scenarios.

I also want the voicing to feel considerate, never overloading too much information into too short of a time, and not making assumptions about what the driver might already know. We found a big win here in giving the system a short-term memory of where you've been, so it can reference that confidently. It can start a departure by saying“Head back the way you came” rather than “Head North”—as you might not know which way north is.

So, creating that “skilled local” feeling really has to be built into the entire system from the ground up, and starts before any of the tone-of-voice considerations.

Did you have any particular hesitations about going down the road of voice technology for this product?

I think two big things have held back the level of functionality in voice: discoverability and latency. Discoverability is about finding out what the system can do. Can I tell it that a road is temporarily blocked? Ask it to repeat something? Tell it to prefer a certain road? LLMs have unlocked some great results here in taking fuzzy inputs from the user and mapping them to structured commands, as well as generating feedback that naturally guides the user towards the available interactions.

Latency is still a challenge during quick back-and-forth exchanges, but the performance in state-of-the-art models is creeping closer and closer to human speed, and I think we're near a tipping point where it becomes a non-issue for most use cases.

Lastly, do you think that voice technology is opening up a field of entirely new products, or it’s more about adding this modality to existing ones? Where do you think it makes the most impact?

From a technical perspective, it's lower bandwidth than screens and keyboards and mice, so it's kind of a slow way to interact with a digital system; I can type faster than I can talk, or scan a screen for the desired information faster than waiting for an entire narration. Voice becomes valuable when using those things is inappropriate for some reason, or if the physical resources they require – your eyes, hands – are unavailable or reduced in function.

Where it becomes really interesting, for me, is when considering the resource of attention. I don't know the science behind it, but I've observed that listening and speaking seems to draw from a different pool of attention, in a way. When doing tasks that are primarily about visual attention and motor control, doing secondary tasks via voice interface seems to work well psychologically. I've used voice in various experiments and personal projects over the years, and one application that has really stuck is the dictation of "distractions” while I'm working.

For example, if it suddenly pops into my head that I need to take out the trash bins for collection, I can just tap a button on my desk and say it out loud, and know that it has been transcribed and will be shown to me when my session is over. Without switching tabs or pulling out my phone, my brain is relieved of whatever itch it had and I'm still focused. I think this is a useful point of view for considering the possibilities of voice: eyes, hands, attention, and smartphones are an unbeatably efficient combo. Voice makes the most impact when, for whatever reason, those resources are not available.