Submitting new species data to be added to classifier

[Tyler-Schwenk]

Hello,

I have been working with The Nature Conservancy for the past few years developing audio monitoring techniques, and in particular training custom classifiers using BirdNET. I have been able to create a solid classifier for the California red-legged frog (Rana Draytonii) which is an endangered species that is being reintroduced to southern California.

I am looking to help get this species added to the standard BirdNET algorithm in order to streamline its use in other services that do not support custom classifiers. Is there a standard way of collaborating on getting a new species added to BirdNET? I would be happy to provide annotated training data or any information about my custom trained classifier.

Any help to point me in the right direction would be much appreciated!

[kahst]

The best way to contribute to BirdNET through donating data is actually to upload recordings to Xeno-canto. We'll regularly pull data from XC and then your recordings will make it into our training data. XC also offers bulk upload if you have a larger set of recordings with metadata (usually a csv with the same columns as the metadata they show on the recording page) if you send them a disk or shared folder link.

[Tyler-Schwenk]

Thanks for the quick reply! I will add my data to Xeno-canto

[Tyler-Schwenk]

@kahst My dataset is growing, and I will be collaborating with a few institutions to combine audio data from multiple regions. I just wanted to clarify if there are any guidelines for how to prepare the data in terms of file length, annotation, or metadata to help ensure it is usable withing BirdNET.

Are there any best practices I should be aware of, separate from the upload guide given by xeno-canto?

[kahst]

Maybe just one tweak: when cropping your recording, make sure the first vocalization of the primary species starts within the first few seconds and the file ends shortly after the last vocalization of the primary species. Sometimes people tend to use only the first and last 3-5 second snippet of XC files due to weak labeling. Other than that, don't apply heavy noise suppression or bandpass filtering to avoid artifacts. At some point, XC will allow bounding boxes on their platform. Labeling background species is crucial, and you might be able to add those boxes later if you already have them.