I am developing neural networks with the aim of detecting, and ultimately identifying, nocturnal flight calls (NFCs) automatically. Neural networks require a lot of sample data: they need audio samples containing targets (i.e. NFCs) and audio samples containing other things that are not targets (all the other stuff we typically ignore, like barking dogs, squeaking bikes, creaking fences, local singing birds etc) so the network can learn to distinguish them. The more recordings I can use from a variety of locations and conditions, the better the algorithms will eventually be. If you are happy to contribute recordings, I’ve detailed below what I am looking for. If you can help with any or all of these, please email me and I will provide details for where you can upload files. Thank you in advance for your interest in this project.
Short clips containing calls identified to species
I need lots (hundreds, if not thousands) of examples of each flight call of each species we encounter in Europe. I need examples of close and distant individuals, so don’t worry if some clips aren’t the best, and if you have a sequence of calls as a bird approaches/moves away, these are great too. I would like calls of ALL nocturnal migrants except Redwing (Turdus iliacus) for which I already have a large sample. If you are only able to send a small number of files, please prioritise the scarcer species as these are the hardest to build up a good sample size.
Key requirements are:
- Ideally audio files should be WAV files but MP3s files are also OK.
- Files should have at least 2 seconds (ideally 5 seconds) of ambient noise before and after the call(s).
- Files can be stereo or mono but should have minimal or no processing. Ideally no noise reduction. The one exception is I think I can accept records with a low cut filter as might be implemented on some recorders to remove rumble (e.g. <250Hz).
- I will need to know date (time if possible), location, your name and the focal species in the clip. If you can put this in the filename like below, that would be very helpful:
- Example file name format: 20190127-0138-Cambridge-Gillings-Turdus_merula.wav.
Short clips containing confusion sounds
If you have saved audio files of things that are confusingly similar to bird calls, or are vocalisations from local stationary birds (e.g. nocturnal chorus), I would gladly receive these too, following the same guidance on file format and naming as above.
Longer audio files
I need longer audio files for two main reasons: 1) to extract samples of ambient background noises to allow the network to learn what isn’t a bird; 2) to test the neural network to check it detects all the calls in a long recording. It would be great to get files that span a range of conditions (e.g. calm to breezy; quiet to noisy neighbourhoods) and a range of bird activity levels, from nights with few birds to busy nights with an abundance and/or diversity of species to test the algorithms work under different conditions. Other requirements:
- WAV files; stereo or mono, with no processing (low cut filter permissible)
- At least 10 minutes duration, but otherwise can be as long as you like
- Date, time, location plus any general indication of content (e.g. was it a night with a strong thrush passage). Ideally following a similar filename convention, e.g. 20190127-2000-Cambridge-Gillings.wav
Audacity labels (OPTIONAL)
Supplying audio files as described above will be great. If you are very keen and want to help even more, the next (OPTIONAL) step is to also supply label files. These are text files, created in Audacity, which mark when in each file each species calls. If this sounds like something you’d like to try, there are some instructions here. I’m happy to talk anyone through this if it looks complicated (it’s not really).