Cutting training clips to < 3 seconds to isolate single species calls

[Geematic]

We're trying to train a classifier for several species of frogs and having issues finding 3 second training recordings.
I'm searching hundreds of hours of recordings and a few species are hard to isolate, many species are frequently calling close together in the same recording- often over each other.
3 seconds is a long time when different species calls are close together.

I understand birdnet pads < 3sec training clips with zeroes.

Can the padded zeroes in the training clips cause issues with training?

Has anyone had any experiences trying this?

[kahst]

Here are a few thoughts that might help:

• you can train custom models with multiple labels per clip; simply name folders "species1;species2". During training, we'll separate the species labels and treat them as independent classes
• we don't expect the padding to be an issue, as long as it doesn't create a consistent pattern (e.g., padding is always exactly 1.5 seconds)
• don't exclusively look for high quality samples for training, training data diversity is key
• if your target species is usually vocalizing among others, your training data should still represent your target domain (i.e., multi-species training samples)
• if you see lots of false positives and suspect there is a pattern, consider creating a "negative" sample folder with name "-species1" which contains all samples that were detected as species1 but were (high-scoring) false positives

[Geematic]

Thanks for this - this will really help us selecting and structuring our training data.