Teaching AI to listen to engines.

Every car person knows the feeling. You start the car, pull out of the driveway, and something sounds off. Not broken, not loud, just different. A tick that wasn't there last week. A hum that rises with the revs. A rattle at idle that disappears once you're moving.

Most of the time you can't describe it well enough to a mechanic, and by the time you get it booked in, the noise has either gone quiet or turned into a bill. EngineEar exists for that gap: the time between "that sounds weird" and "that's going to be expensive".

This post explains how EngineEar listens, what it's good at, and where its limits are.

Your phone is a better microphone than you think

Modern phones have surprisingly capable microphones and enough processing power to handle serious audio work. What they don't have is context. A recording made in a windy parking lot, with the radio on and the phone pointed at the wrong part of the bay, is noise, not data.

So the first thing we built wasn't an AI model. It was the guided scan.

When you start a scan, EngineEar asks what you want to check and walks you through it:

• General health check: a baseline recording across idle and light revs.
• Idle: the steady state, where ticks, misfires and rough running are easiest to hear.
• Cold start: the first seconds after start-up, when some problems are loudest before oil pressure and temperature settle.
• Rev: a controlled rise and fall in RPM to hear noises that track engine speed.

Each flow tells you where to hold the phone, how long to record, and what to do with the throttle. A consistent recording is worth more than a long one, and consistency is what makes it possible to compare scans over time.

From sound to signal

Raw audio is a wave. On its own it isn't very useful to a model. EngineEar converts each recording into a representation of how the sound's frequencies change over time, which is roughly how an experienced mechanic listens: not "is it loud?" but "what's the rhythm, what's the pitch, and does it change with RPM?"

A lot of engine faults have a signature:

• Something that repeats once per combustion cycle behaves differently from something that repeats once per crank revolution.
• A noise that rises in pitch with the revs is tied to rotating parts. One that stays constant probably isn't.
• An irregular stumble in an otherwise even rhythm points somewhere different from a steady, sharp tick.

The models look for patterns like these and compare them against patterns linked to common problems.

Anomaly detection: learning what your engine sounds like

Here's the part we're most excited about. Engines are all different. A flat-plane V8, a diesel four-cylinder and a rotary sound nothing alike, and even two identical cars can sound different depending on exhaust, mileage and maintenance.

That's why EngineEar keeps a scan history for every vehicle in your garage. Your first scans become a baseline. Later scans are compared not only against general patterns, but against how your own engine sounded before. A change from your baseline is often more meaningful than any single recording on its own.

In practice that means scanning when the car is healthy is useful, not wasted. The best time to make your first recording is before you need it.

Reports you can actually read

After a scan, you get a health report: an overall status, what was detected, and what we'd suggest doing next. We've worked hard to keep the language plain. "Possible valve train noise at idle, more noticeable when cold. Worth checking oil level and having a mechanic listen" is more useful than a confidence score and a spectrogram.

Reports are shareable, so you can send one to your mechanic or a friend who knows the platform.

Built for real drivers

We didn't build EngineEar for workshops with calibrated microphones and quiet rooms. We built it for driveways, street parking and the parking lot at work. That shaped a lot of decisions:

• Scans are short, so you'll actually do them.
• The app tells you when a recording isn't usable (too much wind, too quiet, too short) instead of producing a confident-sounding result from bad audio.
• Your garage holds every car you own, from the daily to the weekend project, each with its own history.

What's next

We're continuing to improve the models, expand the kinds of issues EngineEar can recognize, and make trends over time easier to read. If you want to follow the research side of this work, we're writing about it in the Lab.

In the meantime, add your car, run a general health check while everything's fine, and let EngineEar learn what normal sounds like. Learn more about EngineEar.