4 ms·

Would it be too difficult to apply transfer learning ( or create from scratch ), an ML model to train a bird-song recogniser? Edit: There are sonograms too ( h

Would it be too difficult to apply transfer learning ( or create from scratch ), an ML model to train a bird-song recogniser? Edit: There are sonograms too ( https://www.xeno-canto.org/sounds/uploaded/ZWAQHOJFLZ/ffts/XC477551-med.png https://www.xeno-canto.org/sounds/uploaded/ZWAQHOJFLZ/ffts/X... ). Could be done via image classification too. Interesting enough to do on a boring weekend, perhaps?

redvenom 6y ago
This is one of those Kaggle contests. They have a database there that you can try your hand it.
cannam 6y ago
> Interesting enough to do on a boring weekend, perhaps? A weekend, or a PhD! The Machine Listening group at QMUL covers bird audio - see Dan Stowell's work for example: http://c4dm.eecs.qmul.ac.uk/people/dans.html http://c4dm.eecs.qmul.ac.uk/people/dans.html Here's a Bird Audio Detection Challenge from 2018: https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13103 https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111... http://dcase.community/challenge2018/task-bird-audio-detection http://dcase.community/challenge2018/task-bird-audio-detecti... I don't know a great deal about bird song myself, but I gather the subject has lots of interesting subtleties, especially when applied to live audio or real-world recordings where it isn't necessarily known where the bird soounds appear or how many birds there are.
siddboots 6y ago
I've had a bit of a go at this, (although I didn't spend all that much time on it). One thing that challenged my expectations was that it seems you need a lot more data than you might think. Lots of species only have a handful of samples on Xeno-Canto, and the field recordings need aggressive processing in order to isolate the relevant parts, which is hard to do that in an automated way without inadvertently destroying lots of your valuable training data.
mongol 6y ago
This has been done. But I can't say how easy it was. Check out the Android app BirdNet
- Wistar 6y ago
  It's very cool. Here's a demo of it on a (desktop) size screen: https://birdnet.cornell.edu/live/ https://birdnet.cornell.edu/live/

redvenom 6y ago

This is one of those Kaggle contests. They have a database there that you can try your hand it.

cannam 6y ago

> Interesting enough to do on a boring weekend, perhaps? A weekend, or a PhD! The Machine Listening group at QMUL covers bird audio - see Dan Stowell's work for example: http://c4dm.eecs.qmul.ac.uk/people/dans.html http://c4dm.eecs.qmul.ac.uk/people/dans.html Here's a Bird Audio Detection Challenge from 2018: https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13103 https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111... http://dcase.community/challenge2018/task-bird-audio-detection http://dcase.community/challenge2018/task-bird-audio-detecti... I don't know a great deal about bird song myself, but I gather the subject has lots of interesting subtleties, especially when applied to live audio or real-world recordings where it isn't necessarily known where the bird soounds appear or how many birds there are.

siddboots 6y ago

I've had a bit of a go at this, (although I didn't spend all that much time on it). One thing that challenged my expectations was that it seems you need a lot more data than you might think. Lots of species only have a handful of samples on Xeno-Canto, and the field recordings need aggressive processing in order to isolate the relevant parts, which is hard to do that in an automated way without inadvertently destroying lots of your valuable training data.

mongol 6y ago

This has been done. But I can't say how easy it was. Check out the Android app BirdNet

Wistar 6y ago
It's very cool. Here's a demo of it on a (desktop) size screen: https://birdnet.cornell.edu/live/ https://birdnet.cornell.edu/live/