Katir, and knowing what you don't know
Most audio recognition asks whether a sound is one of the classes it was trained on. A room is not made of those classes. Katir is built around admitting that.
Most audio recognition asks whether a sound is one of the classes it was trained on. A room is not made of those classes. Katir is built around admitting that.
Katir is an acoustic monitor that runs entirely on one Android phone. It listens to a room, decides for itself when something happened, and puts anything it cannot place into a queue for a human. No audio leaves the device.
A conventional audio classifier asks: is this one of the 527 things I was trained on? That is a good question if your world is made of those 527 things.
A room is not. A room has a particular fridge, one specific door that sticks, a tap that drips at a rate you would recognise instantly and no dataset has ever labelled, and a cough at three in the morning. None of these are classes anyone shipped. Ask a fixed classifier about them and it will confidently return the nearest label it owns, which is worse than saying nothing.
Katir never assumes it knows what a sound is. It notices that something happened, learns which of those somethings keep recurring, and asks you about the rest.
You name the fridge once. From then on, the fridge is the fridge.
Sounds that repeat get recognised as the same thing without you labelling every occurrence. Sounds the system cannot place are not forced into a category to make the numbers look tidy — they go into a ranked review queue, each item carrying the reason it ended up there. The queue is the product: it is the system telling you where its own judgement ran out.
Capture, detection, recognition and the dashboard all run on the device. There is no account to sign into and no telemetry. The only thing ever fetched over the network is the model itself, once.
Speech is discarded before it is written to disk — not deleted afterwards. The system still records that it heard something and dropped it, so you can audit the gate working without anything reconstructable being kept.
An ordinary phone is enough. This is not a device that needs a hub, a subscription, or a server in someone else’s building.
Scored on a 600-second household soundscape with exact ground truth, at a half-second onset tolerance.
| What | Score | Reading |
|---|---|---|
| Detection accuracy (F1) | 0.86 | against 0.52 for the standard few-shot baseline |
| Short events (F1) | 0.92 | cough, door lock, footstep, light switch |
| Cluster purity | 0.77 | recurring sounds grouped correctly |
| Speed, on the phone | 41× | faster than realtime, full pipeline on-device |
Katir is Android v0.1.0 and in development. More at katir.resonator.am.