Bird Audio Detection: baseline tests and the problem of generalisation | Hacker News Reader

Bird Audio Detection: baseline tests and the problem of generalisation

machine-listening.eecs.qmul.ac.uk

24 pointsjarmitage4 comments

I've been working on matching recorded bird audio from a device to known matches and it's not an easy problem (personal hobby project). Unlike music, birds have dialects within the same species and that can make it harder to match the calls of one to another. Dialects that are different enough that birds of the same species living in the city cannot understand ones that live in the country.

It's an interesting problem, but there's quite a bit more to it than one would think. For some birds like the crow, their calls are complex to the point that they're more like a language. I'd compare it to dolphin communication.

I'm interested in your project. Do you have any blog/site/twitter or what not that I can follow for updates?

If I get something semi usable, I'll probably post it on "show hn". I have a twitter, but I stopped actively using it a while ago[1]. If you're interested, I'll try to post updates there when I have something to update on. My public email is in my hn profile as well.

I do a lot of running and I always hear birds in the forest, but I rarely see them. Given that it's pretty hard to figure out a bird by its song without knowing it ahead of time, I started to wonder if there was a way to match their calls against known samples. As an indirect result of working on the project, I've gotten better at personally identifying their calls, but would still rather automate it through an application, because that's more fun :)

Still very much in a prototype stage and I work on it in spurts in between my day job and other hobby projects. I've been more focused on if I can accurately analyze and match a couple common species to the area, but I should probably step back and see if the typical cell phone microphone can even pick up bird calls from a range of a couple 100 feet. I'm not as worried about that I guess, since it's for a niche audience and I'd find a way to use it myself even if it didn't work quite so well on a phone.

[1] https://twitter.com/yarlyyyy

A friend of mine asked me if it was possible to use machine learning to identify bird calls from recordings. My initial response is: this sound hard & I don’t think the available datasets of identified bird recordings are anywhere near big enough to successfully train a DNN. I’d love to be proved wrong but if the state of the art is 85% accuracy in just recognising bird song in the first place then identifying individual species is a long way off.

Reply on news.ycombinator.com