Unsuccessful at Training a Custom Classifier for birds in eastern Colorado

[ungulate72]

I am attempting to fit a custom classifier for 12 bird species in eastern Colorado, plus prairie dogs using the Birdnet Analyzer GUI, I first trained a classifier using all available clips for each species, and the predictions on test data had >50% error rate for predictions with a confidence score of 1. I trimmed the training dataset down to just good quality recordings (5 to 50 clips per species), refit the custom classifier, and then made predictions on the TRAINING dataset. Both times I also include a Noise category with clips of background noises. It is still getting >50% error for predictions with confidence score of 1. I have watched and followed the online video by Stefan Kahl on creating custom classifiers, but clearly must be doing something fundamentally wrong. Does the training dataset need to have very high quality recordings?

[Josef-Haupt]

Hard to diagnose, but could be a lot of things. Check our short guide on data collection and preparation.