RSPB

Contact: Steffen Oppel <Steffen.Oppel@rspb.org.uk>

Suggestion:

I would be interested to suggest a project as a client, that would allow wildlife managers to count bird calls automatically from acoustic recordings. Although there are several commercial packages available, the ones we have tried so far (Raven, Songscope, SoundID) are unsatisfactory (very high commission and omission errors) and all require substantial up-front investment in defining the variability of calls of the target species.

Ideally, we would like a simple-to-use software tool (R package or GUI) that would allow users to upload recording files (in batches, as there will generally be hundreds), click ‘run’, and then download a table of the number of calls of different species in each recording file.

There has been huge progress in the automatic detection of bird vocalisations in recent years (http://dcase.community/challenge2020/), but most of these high-end algorithms are not accessible to wildlife managers who generally do not have a programming background. So essentially we would be asking students to translate existing algorithms into an accessible platform that would work for wildlife managers in the field. We have many hours of recording from seabird colonies that can be used as training or test data.

Initial feedback:

Thanks very much for this suggestion, and I’m sure it would be an area of interest to students.

I’m fairly familiar with this general field, having recently supervised a PhD that set out to improve machine learning tools for bird call identification by citizen scientists. Based on that experience, my understanding is that the main problem holding the field back is lack of segmented and labelled training data that has been collected in the same environmental context where the system is to be applied (the question of how a model trained with bird calls recorded in captivity can transfer to natural environment is probably one for researchers).

I presume that your colony recordings haven’t been segmented, but it might be interesting to train a model that aims to estimate colony size from acoustic characteristics, rather than counting and classifying individual calls. Do you have a range of recordings from colonies of different sizes? If there are a mix of species in a given colony, are estimates of the relative proportions available in your training data?

My PhD student ended up creating tools for field recording and segmentation of bird calls with a mobile phone, and a game-based interface to educate and practice identification (neither particularly valuable in scientific terms, I’m afraid).