Your skills are needed
Are you an experienced birder able to reliably ID birds by sound?
The BTO has reached out to the West Midland Bird Club for help with a project to review the accuracy of BirdNet, a key tool used to identify bird sounds. To do this, experienced birders able to reliably ID birds by sound are needed to review sound clips and identify them so that the BirdNet results can be evaluated and ultimately improved.
The same technology behind BirdNet is used in the well known Merlin App that is so popular in the birding community and while it’s useful, we know it isn’t always right, so this project to improve the technology will ultimately help us all.
The project is being run by a William Gough, a PhD researcher at the University of East Anglia who is working with the Breeding Bird Survey team (‘BBS’) at the BTO. Here’s what he has to say about the project and how you could take part:
Over the past three years, the University of East Anglia (UEA), in collaboration with the BTO, has been running the UEA Soundscapes Monitoring Project, using acoustic recording devices to investigate soundscapes across BBS squares throughout Britain.
With the support of BBS volunteers, over 60 recording devices were deployed across a range of habitats and landscapes, collecting thousands of hours of audio for future bioacoustic and ecoacoustic research.
These recordings have since been processed using BirdNET, an automated bird sound identification tool that uses machine learning – a form of artificial intelligence – to identify species from their vocalisations. You will have heard before from the BBS team about Merlin and other 'hand-held' auto-ID tools. BirdNET is very similar, only it is a tool use mainly for 'batch' processing. BTO have their own such tool, the Acoustic Pipeline, which can be applied for certain assemblages of species.
Automated identification tools like Merlin and BirdNET are not always correct, and their performance can vary between species, habitats, seasons and times of day. As part of my PhD research, I am investigating how accurately BirdNET performs under these different conditions. To do this, we need to manually validate a subset of BirdNET's identifications by listening to short audio clips and identifying the bird species present, and comparing these to BirdNET outputs.
How can you help?
We have developed an online platform called Audio Checker, which allows volunteers to contribute to this research by listening to five-second audio clips and identifying the bird species they can hear. These identifications will then be compared with BirdNET's predictions to help us assess its accuracy.
Not every clip will contain a clear or easily identifiable bird vocalisation. Some may feature faint calls, multiple species or background noise. Even if you are uncertain about an identification, your response is still valuable.
We would encourage volunteers to return regularly for short listening sessions rather than completing large numbers of recordings in one sitting. As a guide, 20 verifications, which is our recommended minimum number of clips per session, might take around 10 minutes to
assess.
In order to take part in the study, you will need to register for an account. Whilst this work is in collaboration with BTO, this is an entirely separate account to your normal myBTO account.
We welcome contributions from anyone with an interest in birds and their vocalisations, regardless of experience. However, as a BBS volunteer, you will be amongst the most skilled at identifying bird song and so your contribution would be extremely welcome indeed. As part of the initial testing of the checker, a number of ornithological groups have been invited to participate.
Every verification helps us build a valuable dataset for understanding the reliability of automated bird sound identification and informing its use in future ecological research. I hope you are able to take part.
