Training several single species classifiers for hard to isolate calls.

[Geematic]

I am training a frog classifier.
We have recordings that we want to use for training. The issue is that most have more than one species calling over each other. Not always the same combinations, but it's been hard to isolate some.
Training a multiple species-classifier I couldn't find a way to do multi-labelling with birdNET (where a 3 sec training recording might have more than one species calling in it ) other than putting the same training clip in both class/species folders.
Doubling up training data in separate class folders feels like it may cause issues.

Has anyone had experience breaking up training from a multiple species classifier into a few independent single-species classifiers?

For a single species classifier - this what I'd planned for the training folder

species1/
    3sec_audio_with_only_species1.wav etc.
    3sec_audiowith_species1_and_species2.wav etc.
   3sec_audiowith_species1_and_species3.wav etc.
  
-species1/
    3sec_audiowith_species2_and_species3.wav etc
    3sec_audio_with_only_species2.wav etc.
    3sec_audio_with_only_species3.wav etc.
  
NOISE/
    3sec_noise.wav etc.

And repeat for other species classifiers.

I was worried it might require too many more training recordings to make it feasible...

Has anyone tried anything like this? Did it work?
Can anyone see any potential issues?
Any help or advice much appreciated.

This is similar to : #638
But I don't necessarily need it as a single model and im more interested in any way of using recordings for training when you mostly have recordings of multi-species calling over each other to select training data from.

[kahst]

We do support multi-label training samples. You can adapt folder names to contain both species, and your training data could look like this:

• species1
• species2
• species3
• species1,species2
• species2,species3
• -species2
• Noise

During training, we'll split labels and treat samples from multi-label folders as regular training samples.

[atbirder]

Thank You Stefan for this clarification. After following BirdNET, now also on CRAN/R, and Chirpity, I dare to comment. I understand, the underlying question is a kind of "complexity" mode, some of us have in mind. Overlaping vocalisations of diferent frequencies. An algorithm incl. frequency as parameter would also be useful imho for chorus /dusk/dawn) analysation, esp. in the Tropics.
Limd regards from Vienna,
Martin
Projekt Titel: Bioakustisches Monitoring von Vogelarten in Wiederbewaldungsflächen des biologischen Korridors La Gamba, Costa Rica [1] Bioakustisches 'Monitoring PamLaGam2025+' ONLINE https://lagamba.at/forschung/projekte/

[Geematic]

Thanks, Stefan - that's great news. We'll try this.

[Geematic]

After we train, we get each of the multi-labelled folders output as new classes for each multi-labelled folder.

eg. we get a two new classes for "species2;species3" and "species1;species2" as well as species1, species2 and species3 as individual classes.

So when we run predictions we'd get "species1;species2" as a prediction AS well as species1, species2, or species3 individually as classes.

Do these multi-labeled training snips contribute to the other individual classes? (ie. do the classes species1 and species2 get trained by this, or do these multi-labeled samples affect only their own classes?)

I may have misunderstood how this works or there is some setting we haven't set correctly.
Thanks in advance for any insights.

Also If we use a folder "-species1;species2;species3", would that be a counter-training samples for the class
"species1;species2;species3"? (or would it also count against species1, species2 and species 3 individually?

[abfleishman]

I am not 100% sure but I thing the folder names are supposed to be comma not semicolon separated.