Audio Positioning System

Put a microphone in each corner of a room. When someone speaks, the sound reaches the nearest mic first and the others a few milliseconds later. Sound travels about 34 cm per millisecond, so those tiny differences are enough to work out where the speaker is standing. The idea came up about ten years ago and sat in the back of my head until now.

The live app shows a floor plan of the room with a heat map of where sounds are likely coming from, a dot for each one it's tracking, and the timing difference measured between every pair of mics.

How to use it

It opens in Simulator mode, a virtual 5 × 4 m room with a mic in each corner and up to three sound sources.

• Drag the sources A, B and C around the plan and watch the located dots follow them.
• Change what each source plays: speech, claps, broadband noise or a steady hum.
• Turn up the wall reflections and mic noise to see where it starts to struggle.
• Tick "Source A walks around the room" to watch it track a moving talker.

Live mics uses a real audio interface. It needs four microphones plugged into one interface that the browser sees as a single four-channel device, with inputs 1 to 4 going clockwise from the top-left corner. Enter the room size and mic positions under Room and mic geometry.

How it works

There's no way to know when a sound was made, only that one mic heard it a bit before another. So it works with differences.

Every 43 ms it takes a snippet from each pair of mics and slides one against the other to find the shift where they line up best. That shift is the delay for that pair. Four mics give six pairs, so six delays. Those are the spiky traces under the plan.

Then it splits the floor into 5 cm squares and asks of each one: if the sound came from here, what would the six delays be? A square where all six traces have a spike in the right place scores high. The heat map is that score, and the dots are the highest-scoring squares.

Two people talking give two spikes per trace and two hot spots on the map. A wrong pairing of spikes might fit one or two mic pairs, but never all six, which is how it avoids seeing people who aren't there.

Where it's at

In the simulator it separates two or three talkers at once and places them to within a few centimetres, even with a fair bit of echo. That's a best case, because the simulated room matches the maths exactly. A real room adds furniture, longer echoes and mics that are only measured to the nearest centimetre or two, so I'm expecting more like 10 to 30 cm.

A steady hum can't be located at all. It repeats every cycle, so it lines up with itself at lots of different delays.

The hardware is next: a four-input audio interface and four measurement mics. After that, a clap-to-calibrate routine and support for eight mics.