How a neural network helped to get acquainted with feathered neighbors at the dacha

One bird starts singing at sunrise and around 10 a.m. next to our little house in the Nara-Fominsky district of the Moscow region. We sit down to breakfast with the whole family, and our feathered neighbor (or neighbor) also calls his family to breakfast. However, by 10 a.m. he may already have lunch. The bird looks like an ordinary city sparrow, but is a little larger. It sings beautifully, but still… well, it is not a nightingale. The latter has a much wider range of sounds, if I may say so (correct me, musicians, if the term is incorrect).

And we also observe about a dozen birds from year to year, which peacefully coexist nearby and fill the space with a cacophony of pleasant sounds. Some, by the way, build nests and raise offspring right under the roof of our house, where they have found protection from bad weather and predators. This year, our house “sheltered” two families. I decided to put an end to disputes and doubts and find an application that accurately recognizes a bird by sound – “chirping”, that is.

In the Play Market I was looking for those that would be made by a team with the participation of professional ornithologists and scientists. And my choice fell on BirdNet.
The application was developed by the Center for Bioacoustics Preservation at Cornell University (USA), and the IT part was developed by the Department of Media Informatics at the Technological University of Chemnitz (Germany). Thus, this is a joint project of two scientific institutions. The application has a Russian-language version, it is downloaded and works in Russia without problems.

How it works?

BirdNet classifies bird sounds using machine learning. As the creators of the service explained, “The workflow for training the Deep Neural Network consisted of collecting large amounts of audio data, pre-processing this data to create visual representations of the sound, augmenting these visualizations, and finally training a complex model architecture with about 27 million trainable parameters.”
You can read more here: https://www.sciencedirect.com/science/article/pii/S1574954121000273
Here is a link to GitHub, where the creators of the service posted the codes:
https://github.com/kahst/BirdNET-Analyzer

So what kind of bird is it that sings constantly in the mornings under our windows? BirdNet found out: it’s a garden warbler.

The neural network also identified the following species among our neighbors:
Black Swift
Field Sparrow
Magpie
Chiffchaff
Crossbill
Wagtail
Black-headed Goldfinch
Garden warbler
Hoodie
Robin

However, this list is not complete. A couple of times an owl flew in at night. But this bird flies so quietly that you understand that this neighbor is also nearby only when the wind from the flapping of its wings blows you.