Audio search engine can identify Australian wildlife by sounds

December 6, 2023 8:48 PM

Never been done before: This audio search engine can identify Australian wildlife by the sounds they make. Australian researchers and Google have launched a first-of-its-kind search engine that analyses and identifies wildlife sounds.
posted by chariot pulled by cassowaries (12 comments total)

Hey, thanks for posting this! I was the lead researcher on this project, so feel free to ask questions!

I also posted a few notes to Projects, along with a couple example searches which you can try out.

Laughing Kookaburra

Pacific Koel

Chiming Wedgebill
posted by kaibutsu at 10:42 PM on December 6, 2023 [11 favorites]

Nifty - so if I understand this correctly, it feels like this builds out even further into the functional space that Cornell Merlin Sound ID occupies in the Americas and Europe.

That's proven pretty handy to me or identifying bird songs and learning who's in my neighborhood. I can only imagine that Cornell is doing the same sort of gathering data from smartphone users to create heat maps that would get a commercial project a heaping does of the side eye.
posted by drewbage1847 at 11:19 PM on December 6, 2023 [3 favorites]

Merlin doesn't really work well in Australia, as they don't have enough training data for Australian species. A broad species classifier for Australia is a desirable thing.

This search tool doesn't actually do identification, per se - given an example audio clip, it tries to find similar examples. When it works well, it helps bring together a lot of training data very quickly - both positive examples, as well as 'hard negative' examples (similar to what you're after, but not). So we can potentially use this to build towards better species classifiers for Australia efficiently.

But beyond species classifiers, folks working with specific species often have more specific questions, like differentiating different call types, discerning adults vs juveniles, or 'am i actually listening to a lyre bird?' High-quality audio search gives leverage for getting at those more subtle classification questions more efficiently.
posted by kaibutsu at 11:37 PM on December 6, 2023 [4 favorites]

I am An Old. I'm also a crap field biologist - and I've been around some good ones so I know my limitations. In my paper-based education days, we were taught how to use a dichotomous key: splitting the world of possible IDs into one of two options, recursively, until you got to Genus and species. Make an error early in the process and you wander down fainter and fainter paths of unreality. Makins Concise Flora didn't have any pictures; just some line-drawings of obscure details as an aid to the text. I gave up, and reverted to a hello trees, hello sky person when outdoors.

These new on-line tools are bri'nt! They may not be 100%, but they are often good enough = fit for general purposes. It will make getting outdoors more rewarding and informative for ordinary folks and that is A Win. And it scales! as more data gets sifted, more precise ID is possible. Chapeau! Kaibatsu.
posted by BobTheScientist at 12:09 AM on December 7, 2023 [3 favorites]

This is so cool, thank you Kaibatsu! I've been getting into birdwatching lately, I think it's important that some of us are keeping track of what they're planning.

Is the ABC quiz broken? I got 3/3, and I'm not good at call identification, one of the reasons I think this is so cool, and it says I'm in the top 0%? Does that mean everyone gets 3/3?
posted by Audreynachrome at 4:34 AM on December 7, 2023

Can it identify the sounds of hoopsnakes and drop bears?
posted by grumpybear69 at 6:12 AM on December 7, 2023 [3 favorites]

Google's plan:
Phase 1: identify the wildlife by the sounds they make.
Phase 2: get the wildlife to buy stuff.
posted by Cardinal Fang at 6:26 AM on December 7, 2023 [2 favorites]

I love it. Also, the algorithm for Lyrebirds is easy: nothing else sings Nokia ringtones and dirt bike noises.
posted by kandinski at 8:21 AM on December 7, 2023 [2 favorites]

Ahh, but can it identify Australian wildlife by the sounds their victims make?
posted by Greg_Ace at 8:26 AM on December 7, 2023 [3 favorites]

↑ ↑ ↑ ↑ huge dataset for that ↑ ↑ ↑ ↑ ↑
posted by lalochezia at 9:09 AM on December 7, 2023 [1 favorite]

Yeah, but they all sound like this.
posted by Greg_Ace at 10:06 AM on December 7, 2023 [1 favorite]

kaibutsu - thanks for the explanation. I love seeing how computers and phones aid with mass data collection and correlation for things likes this (as opposed to the money extracting varieties of consumer data mining).
posted by drewbage1847 at 12:16 PM on December 7, 2023