Problems while creating custom classifiers

[ankithajayanth]

Hi, I am trying to create a custom classifier for birds found in urban areas with a lot of noise in a west Indian city. I have around 20 species for now with ~100 recordings each as well as some samples for noise which includes squirrels, traffic and other noise.
I tried training the classifier and when perform batch analysis with it, all the noise gets very confidently identified as all the species. I am wondering that this problem could be because birds that have contact calls of around 2-3 calls in the 3 seconds also have noise in the samples and that is why even the noise gets picked up as a bird call?

For example an Oriental magpie robin's calls look like this:

So does the classifier also pick the noise in between and then try to correlate it with the robin's call?

[tphakala]

Please see this issue #791, version 1.5.0 is last version custom classifier creation works on, on any version above that it behaves exactly as you describe.

But to answer your question is noise between calls being learned as well, yes it is, you need to isolate those noise parts and feed them to either background class or species negative class.

[ankithajayanth]

Thanks @tphakala , So you would suggest I run it on v1.5.0 and see if I get anything different?

[tphakala]

If you want a working classifier you must use v1.5.0. If classifier built with that version works well enough then you have sufficient noise data in background class already, if not then start isolating noise from species learning data.