Why Nche can’t hear your birds yet.

The idea.

A sick flock often sounds sick before it looks sick. A few birds coughing. Sneezing that wasn’t there last week. A wet rattle in the breathing, worse at night and in harmattan. A good farmer hears it. A tired farmer at 6 am might not.

So we set out to build Nche’s next check: hold your phone up in the house for ten seconds, and Nche tells you if it hears signs of a breathing problem, and when to call your vet.

What we tried.

A listening model learns from recordings: thousands of them, from healthy houses and sick ones. Researchers have shared recordings from farms abroad, and from one study at a Nigerian university. We trained on those.

Before any farmer heard a single answer, we wrote down a test it had to pass: on Nigerian birds it had never heard, catch most sick flocks, rarely cry wolf, and give a real answer most of the time, not a shrug. We wrote it down before training, so a good-looking result couldn’t talk us into moving it.

What happened.

On farms like the ones it learned from, it did well. On Nigerian birds, it fell apart. Half the time it wasn’t sure. It missed too many sick flocks.

A check that’s wrong half the time is worse than no check at all. It teaches you to ignore it.

Four tries in two days.

1. As it was. Good on farms abroad, poor on Nigerian birds. That was the first surprise.
2. Every recording heard a hundred different ways. We made the model hear each clip in other buildings, through cheap and good phones, over the hum of a generator, a fan and rain on a zinc roof. It got better on Nigerian birds, the first real progress. But it got jumpy too, and started hearing trouble in healthy flocks.
3. A gentler version of the same. Steadier, fewer false alarms. Still not good enough to pass.
4. Teaching it what a Nigerian farm sounds like. We gave it real background sounds from a Nigerian farm (people talking, feeding, pecking) and told it “this is just noise”. It backfired: it learned that anything that sounded like that farm was noise, birds included.

None of the four passed. But each one told us something, and the last one told us the most.

The surprising part.

Again and again, the models weren’t learning what a cough sounds like. They were learning where a recording was made: the microphone, the walls, the hum of a particular house.

In one set of recordings, the healthy birds and the sick birds had been recorded in different rooms. A computer could tell them apart from the background alone, in the quiet moments, with no bird making a sound. A model that learns the room learns nothing about your farm.

That’s not anyone’s fault. Research recordings are made for research. But it means the only honest way to teach Nche what a Nigerian poultry house sounds like is to listen to Nigerian poultry houses: lots of them, healthy and sick, recorded the same way your phone will record yours.

So it isn’t out yet.

Listening stays coming in the app. You can already record your birds and the app keeps the recording, but it won’t pretend to judge it until it passes that test on Nigerian farms.

The photo check is unchanged: a photo of fresh droppings, a first look in seconds for signs of coccidiosis, Newcastle disease and salmonellosis, right about 8 times in 10. Your vet makes the call, always.

What we’re doing instead.

• Recording partners. We’re paying farms to record their houses with the Nche app for two weeks: ten seconds a house, three times a day. Healthy flocks and sick ones, from many farms, so the farm can’t give the answer away.
• Vets confirm what’s wrong. When a flock has a breathing problem, a vet looks at it and confirms what it is. That’s what makes a recording mean something.
• Your privacy first. People’s voices are taken out before any recording is used, and no farm is ever named.
• The same test, no shortcuts. When there are enough recordings, the next model faces the test we wrote down before training. It goes live only if it passes.

It’s slower. It’s the only way it ever works for farms here. If you’d like one of the paid places, or you’re a vet with a client whose birds are coughing, here’s how.

Building here.

I run a commercial poultry farm in Nigeria. Nche is the tool I wished I had on my own farm, and a few birds starting to cough is exactly the warning I wanted it to catch. This week was a reminder of the unglamorous part of building AI for African farms: the data you need often doesn’t exist yet. So you go and make it, farm by farm.

We’ll tell you when it passes. Until then, it stays “coming”.

Chinedu OkereFounder, Nche