Nocmigging, the study of nocturnal bird migration, is increasingly popular but requires humans to analyse recordings on software such as Audacity, which is very time-consuming. However, artificial intelligence could do the work for us! To start with, bird detection could save a lot of time by extracting sections of recordings with bird calls in them, only needing humans to identify birds from the calls. After that, the code could be further developed to extend to bird call identification as well, which would allow autonomous recording stations to be created.
I found this by coincidence - I have been thinking about the same kind of system, using AI to extract bird calls for users to identify. Do you have something practical in mind, hardware or software?
I have read some articles about using AI to identify species from nocmig recordings, but have not seen any tools that could pick out (almost) any bird call, but without the need to identify the species. I've been planning to try this out, using some of my own recordings. I'm currently building a tool to annotate spectrograms, so that training AI could be tested with those. There are some open datasets also, but I have not checked those first-hand, nor I have much experience on using AI (yet).
PS: BirdNET app (Android) is pretty good in identifying many birds from their songs: https://birdnet.cornell.edu/
Sorry for the delayed response, I somehow missed the email notifying me of your comment.
I haven't researched the project too much, as I'm very busy at the moment but first I aim to get the software working on a normal computer with recordings downloaded from a recorder. Eventually, my aim is to have the software run 'on-device', with data (recordings, species and times etc.) uploaded for the user to view. This would enable autonomous nocmig stations to be created.
Still building an automated system to create 10-second audio segments & spectrograms out of long recordings, and a tool to annotate / validate the segments. This is also a project of learning Python for me, so progress is not that fast. Source: https://github.com/mikkohei13/Loxpyt
For some reason a part of my latest message went missing...
Meant to say that I tried Birdvoxdetect with few recordings, but results were mixed: It recognized many faint bird calls very well, but missed majority of the calls altogether, even clear ones (e.g. Redwings). Probably due to it being trained with North American species?
I hope that it doesn't recognise some calls because it only recognises North American species (although this doesn't seemed to be clearly stated anywhere in the documentation), because with a bit of retraining for European species, it could be a good ready-made option. If this is the case, I'll use birdvoxdetect and focus on making good hardware for it to run on in the field for my project. I'll have to try it out when I have the time.
Is there something that can generically identify birds by their songs. Many times I've been out hiking and wondered what bird I had heard. Trying to search for it later online is kind of tough. It would be interesting if there were a machine-learning app that could do it live in the field. Preferably away from Internet connections.
I'd never heard of this. I'd always thought most birds were pretty quiet and settled down after dark. Do only certain species make sounds at night? Are they songs or is it sounds associated with flying?
Many birds are silent and settled down after dark but some species are particularly known for nocturnal movements. Some of the birds recorded are probably local, such as a Moorhen moving from one pond to another, but many are long distance migrants. For example, recently many UK nocmiggers have recorded Redwings and Fieldfares which breed in places like Scandinavia but have now come to the UK for the winter. However, it is thought many species that fly over are silent so many birds are undoubtedly missed.
The birds flying over are all doing their flight calls, so these are what the project aims to identify.
I found this by coincidence - I have been thinking about the same kind of system, using AI to extract bird calls for users to identify. Do you have something practical in mind, hardware or software?
I have read some articles about using AI to identify species from nocmig recordings, but have not seen any tools that could pick out (almost) any bird call, but without the need to identify the species. I've been planning to try this out, using some of my own recordings. I'm currently building a tool to annotate spectrograms, so that training AI could be tested with those. There are some open datasets also, but I have not checked those first-hand, nor I have much experience on using AI (yet).
PS: BirdNET app (Android) is pretty good in identifying many birds from their songs: https://birdnet.cornell.edu/