Unsuccessful at training a classifier for Koala bellows

[LaraDani98]

I am trying to create a classifier for koala bellows. I have a set of training data of around 2.700 3-second snippets of koala calls of different quality, intensity, and range, and another ca. 9.000 snippets of common sounds that could be mistaken for koalas (e.g. cows, dogs, windfarms, traffic, weather, pigs, etc.). Koalas call at a very low frequency of a maximum of 8000 Hz, so that is also the range limit I give the identifier before training it.

When I plug this data into the program, and test it on two test files afterwards, it keeps giving me hundreds of false positive koala sounds. It mistakes everything from cows to wind to just background noise as a koala, even though I have already gone over my koala data and removed all sounds that are very low quality, or just low volume so they barely show up on the spectrogram. The model that only distinguishes between noise and koala is even less successful than the one that distinguishes between koalas and different sounds.

I would appreciate any help and ideas that you might have as to what could be causing this, and how I could improve the model. I am guessing that the low frequency could be an issue, as this program was developed with birds in mind. Is there a way to fix this issue, some kind of trick or a setting I could try out? If anyone has had similar issues, please share your experience with me!

All the best, Lara