Introduction
Hey, I’m Kamaleswar. I’m a master’s student in Computer Science at Blekinge Institute of Technology in Sweden. And for the past few months, I’ve been listening to birds. Well, not me personally – my AI has.
Here’s the thing. On a small island called Stora Karlsö in the Baltic Sea, researchers have been recording guillemots (a type of seabird) for years. They have over 12,000 hours of audio. Twelve thousand. That’s like listening to the same playlist for a year and a half straight.
And most of it has never been analysed.
Why? Because you can’t just sit and listen to 12,000 hours. You’d go insane. So ecologists only look at tiny samples, maybe 1-2% of the data. That means rare stuff like weird calls, distress signals, maybe even completely unknown vocalisations just sit there, unheard.
That’s where I come in.
The problem
Guillemots are kind of a big deal. They’re like the canary in the coal mine for the Baltic Sea. Their health tells you about fish stocks, pollution, climate change, all that stuff.
But if you can’t listen to their calls properly, you’re missing half the story.
I didn’t want to build another "bird identifier" – those are cool, but they only tell you what species is calling. They don’t find things nobody has ever heard before.
I wanted to build something that finds the unknown.
What I’m actually building
So here’s the plan.
I take the audio and turn it into spectrograms, those colourful time-frequency pictures you see in bioacoustics papers. Then I train an autoencoder (a type of neural network) to learn what “normal” colony sounds look like. Wind, waves, background chatter, the usual.
When the autoencoder hears something it can’t reconstruct well – that’s an anomaly. A weird sound. A potential rare call.
Then I use active learning. Basically, the AI doesn’t just dump all anomalies on a poor ecologist. It picks the most interesting ones – the ones it’s most confused about – and asks for human feedback. Over time, it builds a vocabulary of call types.
The goal? A system that can scan thousands of hours, flag only the really unusual stuff, and help researchers discover new vocalisations without wasting years of listening.
How the RS funding helps
This is where the hardware comes in.
I have access to the existing recordings, but they don’t come with behavioural notes. I don’t know if a certain call happened when a bird was alarmed, or feeding, or just chilling.
To validate my AI, I need to record new audio while actually watching the birds. That means going to Stora Karlsö, setting up microphones, and noting down behaviour.
So I used the RS Student Project Fund to get proper field recording gear. Here’s what I picked:
- Zoom F6 field recorder – 32-bit float recording means I don’t have to stress about gain staging. Quiet chick calls and loud alarm calls can both be captured without clipping. Also has 6 channels, so I can place mics at different nests.
- Sennheiser MKH 8020 microphones – omnidirectional, flat frequency response, super low self-noise. Perfect for capturing natural bird calls without colouring the sound.
- Rode Wireless GO II – lets me keep the recorder far away from the colony so birds aren’t disturbed. The mic can be 20-30 meters from the recorder.
- Windshields – Stora Karlsö is windy. Very windy. Without proper wind protection, recordings are useless.
This gear lets me create a verified reference library. When I see a bird do something – alarm, begging, whatever – I have a clean recording of that exact behaviour. That becomes ground truth for my AI.
Conclusion
This project isn’t about building the perfect AI. It’s about building something that actually helps researchers deal with massive datasets.
I’m not going to claim I solved overlapping calls or discovered 10 new call types. But I did build a working pipeline that finds candidate events, clusters them, and prioritises them for review. And with the RS funding, I can now validate those findings in the field.
If you’re a student working on a similar project, just start. You’ll mess up, you’ll get confused, but you’ll learn way more than from any textbook.
And if you ever have to listen to 12,000 hours of seabird audio… don’t. Let the AI do it.
Comments