Merlin Bird ID gives your child an ear for birdsong

Lay a phone flat on a wall on a September morning, open Merlin Bird ID and tap the microphone. Within seconds a list builds itself: a name, then another underneath it, each one appearing the moment the app hears something it recognises. Children seeing this for the first time almost always do the same thing. They look up from the screen, scan the branches, and ask: where?

Nothing has changed outside. The garden was always that loud. The child just had no way in.

The sense nobody gets round to learning

Birding by ear is one of those skills with a brutal entry fee. You have to hold a few hundred songs in your head, then pull one of them out of a hedge full of traffic and wind and other birds. Most people who would enjoy it never start. Merlin, which is free and comes from the Cornell Lab of Ornithology, drops that entry fee to about ten seconds.

What it does under the bonnet is stranger than it looks. While the microphone records, the app turns the audio into a spectrogram, a picture of which frequencies are present and when. That picture is fed into what Cornell describes as “a modern computer vision model called a deep convolutional neural network”, the same family of model used to recognise objects in photographs.

So Merlin is not really listening. It is looking at a drawing of a sound. That is a lovely thing to explain to an eight-year-old, and a good early lesson in how rarely a machine is doing the thing you assumed it was doing.

The training is worth knowing about too. Cornell’s engineers used 140 hours of audio containing bird sounds plus another 126 hours of non-bird background sounds, and experts from the Macaulay Library and the eBird community marked the precise moment each species called rather than labelling a whole recording. That fine-grained work is what stops the model falling apart when several species sing at once.

Two practical notes for parents. Merlin works offline, so a park with no signal is fine. And Cornell states plainly that recordings “are saved only on your phone and are never transmitted automatically to the Cornell Lab”. Nothing leaves the device unless someone chooses to send it.

How good is it, really

Good enough to be worth your Saturday, not good enough to be an oracle. A study in Ornithological Applications, published in August 2025, ran 144 paired point counts against human observers in a spruce-fir forest at Winter Harbor, Maine. The humans logged 382 individual detections to Merlin’s 222, with precision at 92% for the humans and 86% for the app. Merlin caught four species the humans missed and also reported twelve that were not there.

Cornell says as much itself, in its own help pages, before anyone else can: “Merlin’s suggestions are just a starting point.” Which is the honest way to hold it. The app opens the door; the child still has to walk out and find the bird.

Last night, a billion of them went over

Here is the part that tends to stop a child mid-sentence.

BirdCast, run by the Cornell Lab with the University of Massachusetts Amherst, Purdue University and the University of Illinois Urbana-Champaign, reads the United States weather radar network with machine learning and converts it into a live count of birds migrating overhead in the dark. It has been mapping live migration since 2018. On the night of 8 October 2025 it counted 1.25 billion birds aloft, the biggest night it has recorded, and the first billion-bird night had come only two years before that.

Anyone can open the site during the season and pull up the estimate for their own county. Not the country. Their county. Which means a child can check, over breakfast, how many birds crossed their own patch of sky while they were asleep, and work out that this has been happening every autumn of their life, silently, a few hundred metres up.

The view from Singapore

Honesty first, because this is where the shine comes off for us. Sound ID currently handles roughly 2,000 species, which Cornell says “covers most birds of the US, Canada, and Europe, and also includes common, widespread birds of India and Central and South America”. The same page adds: “If you are outside of these regions, Merlin will not be able to identify most birds near you.” Southeast Asia is still a gap. Point it at a koel and it may well shrug.

The season, though, is not a gap. NParks advises visiting Sungei Buloh Wetland Reserve between August and April for migratory shorebirds, among them the Whimbrel, Common Greenshank, Common Redshank, Pacific Golden Plover and Lesser Sand Plover. The reserve sits on the East Asian Australasian Flyway, one of the world’s largest, and its mudflats work as a refuelling point for birds passing through. So our own migration is underway right now; the app simply cannot name most of it yet. Merlin’s browsable guide and recordings for more than 10,000 species worldwide still make a reasonable field guide in a pocket, and it is a very good thing to install before a trip to Europe, India or North America.

If the wider business of machines interpreting animals appeals to your child, we wrote about AI learning animal languages in August.

What it actually gives them

Not a bird list. A recalibration.

A child who finds out how many species are in the tree outside the kitchen, and that a river of birds crosses the sky above the house on autumn nights, has had the size of their world quietly adjusted. That used to take a decade of patient listening, or a research career. This month it takes a phone left face up on a wall for ten minutes.

Go and leave one there.