How do you deal with songs vs. calls when assessing BirdNET detection rates?

Question

I am using BirdNET in projects to assess breeding bird acoustic phenology. BirdNET has provided a compelling opportunity to transition away from acoustic indices (which, though useful, provide challenges for easy interpretation) and toward characterizing species-based activity levels instead.

One of the challenges (and opportunities) of BirdNET is that it gives a confidence score output for the species of interest, but it does not classify the detection as a song or call. In avian species, songs and calls are largely believed to have different functions. The song/call issue also presents challenges in verification, particularly if a verifier knows the song well but is less familiar with the various call types a bird may have (I've worked on developing a flexible / generalizable solution to this problem here, but it's not perfect).

In my project, I would prefer to focus more on songs, and am having trouble figuring out how to deal with the songs vs. calls BirdNET detects. I suppose one solution could be to build an additional classifier on top of BirdNET that classifies further into a song or call, and another might be to estimate song vs. call rates throughout the breeding season based on a verified subsample.

How have others dealt with this?

Answer 1

I would recommend training a model from scratch to specifically recognize the classes you are interested in (rather than training a model to perform a second detection step). This is preferable because if you use two sequential models, your model's performance will be dependant on on high recall (low false-negative rate) of BirdNET.

You could take a "fine tuning" approach by starting with a pre-trained model (on other classes), which can reduce the training data needed to train a good model.

Note that many bird species have several vocalization types, so "song" vs "call" may be a simplification of the diversity of vocalizations. You'll need to decide on a set of classes that you want to identify for any given species.

Answer 2

I understand that you are also trying to sort by species, which was not an emphasis of my research, but I am working in an ecosystem that isn't yet well supported species-wise on BirdNET. I created a custom classifier to sort animals into different categories, and my avian categories broke down different types of bird vocalizations (song, call, trill, etc) and it worked well. Not sure if that's helpful for your particular situation, but just wanted to share in case it's helpful.