What if drivers could just talk to the map?

A sea of stuff happening. Almost none of it reported.

Millions of trips happen daily across Southeast Asia. Drivers see flooded roads, sudden closures, surprise checkpoints—the kind of info that could save hundreds of people a detour.

But the old reporting flow asked for this, at 60 km/h:

Tap Report

→

Pick category

→

Pick sub-category

→

Generic "Thanks"

Most people just… didn't bother.

Same app. Very different worlds.

All drivers look the same in a database. On the ground, it's a different story—and designing one voice experience for all of them was the first trap to avoid.

The 4-wheel drivers

Enclosed cabin. Quieter environment. Comfortable with voice—likes looking more "professional" with hands-free tech.

Voice-friendly Silent with passengers onboard

The 2-wheel bike riders

Phone on handlebar. Wind and traffic noise. In Vietnam, Bluetooth earpieces aren't even legal while riding.

Noisy environment Far from mic High reporting need

Across both groups, one thing was consistent: drivers genuinely want to help each other. In Jakarta, drivers would go out of their way to warn others about flooding—they just needed us to make it worth their effort.

We broke things, watched them fail, adjusted, repeated.

The hackathon prototype proved drivers loved the idea of "just talk what you see." The execution took a few rounds.

01

OS-level speech + snackbar confirmation

Hackathon prototype. Fast to build, but the install barrier killed it before drivers ever heard a word.

~90% failure rate Required package install Snackbar unreadable at speed

02

Redesigned confirmation sheet + integrated transcription + scripted follow-up

Pulled transcription into the app and replaced the snackbar with a large glanceable sheet — but scripted follow-ups hit a ceiling.

No install friction Readable at a glance Richer incident data Stiff and templated Didn't adapt to driver's words

03

Shipped

Real-time AI conversation + smart follow-ups

The AI listens to what the driver said and responds in kind. No scripts. No menus.

Contextual follow-ups Majority completed multi-turn Feels like a conversation

Every "yes" had a "but."

• — Voice is not universally better. It's contextual. With a passenger, many drivers go quiet. In Vietnam, motorbike riders can't legally use earpieces. So voice became a choice, not the default—manual reporting still lives alongside it.
• — Richer data = more driver attention. Each follow-up question costs cognitive load. We kept it to the smallest useful set—closer to radio chatter than a call centre script.
• — Reports flow through validation before changing a route. A single voice report isn't enough to reroute thousands of people—driver votes, image analysis, and passability checks all run first. Running this pipeline for millions of drivers adds up, so we modelled per-report costs with the data team and explored volume-based options. Designing the feedback so drivers know their report "did something"—without exposing that complexity—was its own design challenge.

Tap once. Talk. Done.

The shipped version replaced every tap with a single voice utterance. But the same pipeline that let drivers report incidents could power almost any action that previously required them to stop and look at their phone. We opened the architecture to other teams. What started as a reporting feature became a foundation.

Voice reporting, end-to-end

Tap the microphone, describe what you see. The AI classifies the incident, asks one or two contextual follow-ups, then confirms in a large glanceable sheet. Under 15 seconds. No menus, no category picking.

Query, don't just report

We extended the same channel to let drivers ask questions, not just submit reports. Say "how's the traffic ahead?" and get a spoken summary without leaving the navigation view. Same interface, same utterance—the map talks back.

Reach passengers without pulling over

Drivers often need to contact passengers mid-trip—a late arrival, a pickup change, a quick question. Working with the Communications team, we wired the voice pipeline into messaging and calls. Say "text passenger I'll be 5 minutes late" or "call passenger." The passenger gets a normal message or call. The driver never touches the screen.

Find parking before you arrive

With the Maps team, we extended voice to parking. Ask for nearby spots on the way to drop-off and get a ranked list—nearest and cheapest—navigated automatically.

Numbers that (honestly) surprised us.

In pilots, the majority of invited drivers actively used the feature—and nearly half completed at least one follow-up question. Strong signal that drivers tolerate short, meaningful back-and-forths when they feel useful.

Flood reporting noise in Singapore dropped dramatically when voice reports were combined with external sources—meaning ops teams could finally trust the signal instead of manually filtering it.

What I'd tell myself at the start.