Audio speed modification works to train model at 10x but not 5x

[dawnbarlow]

Hi BirdNET team,
I am training a model to detect low-frequency baleen whale calls. I am therefore using the Audio speed modification feature in the Train and Batch Analysis tabs, adjusting the minimum and maximum bandpass frequencies accordingly. At 10x speed, using a bandpass frequency range of 30-120 Hz (entered as 300-1200 Hz), the model seems to learn well and the training curves look good. When I use 5x speed (150-600 Hz bandpass frequency range), which is what I'd ideally like to use, the model stops after ~12 epochs and the training output plot produces a flat line. Any thoughts on whether this is an issue on my end with how I'm inputting data for training, or whether the base BirdNET model is not sufficiently trained for such low-frequency signals, or whether it is a bug with the GUI?
Thanks!
Cheers,
Dawn

[Josef-Haupt]

Hey Dawn! Could you provide us with some of the training data?

[dawnbarlow]

Hi Josef,
Here is a .zip file containing clips with the various training classes. The D calls are the target signal that we want to reliably detect and classify in this case.
[https://oregonstate.box.com/s/z5f9tr8g7dpsw3424dbxbn5ojcw7u304]
Thanks for any feedback!
Cheers,
Dawn

[Josef-Haupt]

Unfortunately we can't access box since we don't have an account at oregonstate uni or a box plan.

[kahst]

I can access box through Cornell, but the link seems to be expired.

[dawnbarlow]

Ah, my apologies. Can you please specify what training data will be most helpful for your troubleshooting? i.e., the training clips, the custom classifier, sound files themselves, or outputs?

[dawnbarlow]

In the meantime, below are screenshots from the issue I'm describing. I just used a mostly different training dataset and ran into the same thing. When I input the below parameters:

I get the following output after just a few epochs.

However, when I use 10x speed, like this:

I get a far more reasonable looking output, like this:

Any insights as to what is causing the discrepancy?

[Josef-Haupt]

We released v2.1.1 which should fix the audio speed problems: https://github.com/birdnet-team/BirdNET-Analyzer/releases/tag/v2.1.1
Please have a look and let us know if you encounter any problems