Kaggle Birdcall Identification Competition
June 15, 2020 12:17 PM Subscribe
Kaggle Birdcall Identification Competition
We just launched a machine learning competition for bird identification in soundscapes! This is a surprisingly difficult problem, which can ultimately help with ecosystem health monitoring (for example, if birds X, Y, Z are present, you can make inferences about their food and predators). I've been building models in this space for a couple of years, working with both the Cornell Lab of Ornithology and Cal Academy of Sciences in my spare time. I'm excited to see what sort of ideas we get from the larger community!
We just launched a machine learning competition for bird identification in soundscapes! This is a surprisingly difficult problem, which can ultimately help with ecosystem health monitoring (for example, if birds X, Y, Z are present, you can make inferences about their food and predators). I've been building models in this space for a couple of years, working with both the Cornell Lab of Ornithology and Cal Academy of Sciences in my spare time. I'm excited to see what sort of ideas we get from the larger community!
Role: coordinator, researcher
Ah, interesting story!
Audio classification has come a long way in the last few years, but this still seems to be a hard one.
The Xeno-Canto website provides lots of crowdsourced audio from thousands of species. Models trained on Xeno-Canto audio do a great job on other (held-out) Xeno-Canto audio. But when we go to soundscapes, the models do much worse. The big difference seems to be 'active' recording: the XC examples are generally very clear examples of a particular bird, while the soundscapes may be a) less obvious examples (the sorts of things a recordist would not know how to label, for example), or b) just muddier, since the microphone isn't pointing right at the bird in question. To deal with (b), we add lots of data augmentation to make the clear songs muddy during training, but the gap persists.
The XC-to-XC score is about 80% in cMAP (a long-to-describe metric), while the XC-to-soundscape score is usually 10-20%.
posted by kaibutsu at 4:11 PM on June 15, 2020 [1 favorite]
Audio classification has come a long way in the last few years, but this still seems to be a hard one.
The Xeno-Canto website provides lots of crowdsourced audio from thousands of species. Models trained on Xeno-Canto audio do a great job on other (held-out) Xeno-Canto audio. But when we go to soundscapes, the models do much worse. The big difference seems to be 'active' recording: the XC examples are generally very clear examples of a particular bird, while the soundscapes may be a) less obvious examples (the sorts of things a recordist would not know how to label, for example), or b) just muddier, since the microphone isn't pointing right at the bird in question. To deal with (b), we add lots of data augmentation to make the clear songs muddy during training, but the gap persists.
The XC-to-XC score is about 80% in cMAP (a long-to-describe metric), while the XC-to-soundscape score is usually 10-20%.
posted by kaibutsu at 4:11 PM on June 15, 2020 [1 favorite]
This is really interesting. I do work for a company that makes bird guide apps. One of my first questions was, "what about Shazam for birds?" and was told some of why it's been such a challenge to achieve so far.
posted by johngoren at 7:25 AM on June 25, 2020
posted by johngoren at 7:25 AM on June 25, 2020
« Older Typewriter Stories with Doctor Sparks... | Break Into Us... Newer »
posted by Going To Maine at 2:41 PM on June 15, 2020