We Built an AI That Listens to Machines — In One Evening

Last night, me and Otto sat down with a hand mixer, a laptop, and a simple question: can a computer learn to hear when something is wrong with a machine?

Ten hours later, we had our answer.

What We Built

We built a working acoustic fault detection system, a program that listens to a machine through a microphone and tells you in real time whether it's running normally or something is off. No sensors, no expensive hardware. Just a mic and some Python.

The interface? A cartoon grandma. Happy and blue when the machine is healthy. Angry and red when something's wrong. It updates live in your browser every few seconds.

How It Works

The core idea is simple. Every machine has an acoustic fingerprint — a unique sound profile when it's running normally. If that sound changes, something has changed in the machine.

We recorded 40 clips of a hand mixer running normally in different positions and distances from the mic. We then trained an Isolation Forest model . an AI that learns what normal looks like and flags anything that doesn't match as a fault. No faulty recordings needed. The model figures out anomalies on its own.

The stack was entirely Python:

• sounddevice — recording audio from the mic

• librosa — converting audio into Mel spectrograms, visual maps of sound

• scikit-learn — the Isolation Forest model

• Flask — serving the live web dashboard

Where This Could Go

Right now it runs locally on one laptop. But the concept scales further than a hand mixer in a student apartment.

The same approach could be applied to industrial motors, HVAC systems, wind turbines, or any machine that runs continuously. Instead of waiting for something to break — which is expensive and sometimes dangerous — you catch the fault the moment the sound profile starts changing.

Next steps for this project would be adding persistent logging so you can track fault history over time, phone notifications when a fault is detected, and eventually deploying it to a small dedicated device like a Raspberry Pi that sits permanently next to the machine.

The grandma stays though. Non-negotiable.

Built in one evening by Anton Heinonen and Otto Oivanen, April 2026