What is BirdAmp?

What it does

You give BirdAmp some audio — recorded live in the app, or a file you already have — and it returns a ranked list of the species it heard, with a timestamp for each detection. Tap a detection and you can play back that moment, boosted, to hear what the model heard.

It is a tool for answering "what was that?" in the field, and for going back through recordings you made without knowing what was in them. Birders use it to confirm a suspicion; people who simply like the noise outside their window use it to find out whose noise it is.

How Neural Boost works

A field recording is rarely one bird in silence. It is one bird, plus three others, plus wind across the microphone, plus a road half a mile away, plus your own coat sleeve. The identification problem is largely a separation problem.

BirdAmp cuts the clip into short overlapping windows and converts each into a spectrogram — a picture of which frequencies are present over time. A neural network trained on a very large library of labelled bird vocalisations scores each window against every species it knows. Boosting a detection re-synthesises just that bird from the mix, which is both how the app shows its working and how a faint song becomes audible enough for you to check.

Scores from overlapping windows are then corroborated against each other, so a species that appears once in a single window is treated more sceptically than one that recurs consistently through the clip.

Why it runs on your device

The whole model ships with the app and runs locally. Nothing is sent away to be identified.

• It works with no signal, which is where a lot of the good birds are.
• There is no round trip to a server, so results arrive as fast as your phone can compute them.
• Your recordings are not collected. What you record stays on your device unless you deliberately share it.

How to read a result

Every detection is labelled High, Medium or Low confidence. Those labels are deliberate: a raw percentage invites you to read a precision the model does not have, and "87%" tells you far less than it appears to.

• High — the model is confident. On a clean recording of a common species, this is usually right.
• Medium — plausible, worth checking. Play the boosted audio and see whether you agree.
• Low — a suggestion, not an answer. Treat it as a prompt to listen again rather than a name to write down.

What it cannot do

Being straight about the limits is more useful than overselling the thing:

• Several birds singing at once will confuse it. It may name the loudest and miss the interesting one behind it.
• Heavy wind on the microphone can swamp a recording beyond recovery. No amount of processing rescues a clip that is mostly rumble.
• Mimics are a genuine problem. A Starling doing a passable Curlew will be identified as what it sounds like, because that is all the audio contains.
• A very distant bird may simply not have enough signal left to work with.
• Similar-sounding species — Marsh and Willow Tit, or Chiffchaff and Willow Warbler at the wrong moment — are exactly where a confident answer deserves the most scepticism.
• It identifies vocalisations. Non-vocal sounds such as a woodpecker drumming are a different problem and are not reliably covered.

Almost all of these improve with a better recording, which is the subject of the recording guide.

Where you can use it

BirdAmp runs in the browser, so you can upload a file and get an answer without installing anything. Native apps for iOS and Android add live recording, a saved library of your identifications, and better performance on long clips.

Read next