Separating the tweet from the chaff
Scientists from Queen Mary University of London have found a successful way of identifying bird sounds from large audio collections, which could be useful for expert and amateur bird-watchers alike, and be a particularly valuable tool for conservation.
The analysis used recordings of individual birds and of dawn choruses to identify characteristics of bird sounds. It took advantage of large datasets of sound recordings provided by the British Library Sound Archive, and online sources such as the Dutch archive called Xeno Canto.
The datasets gathered in the field by many sound recordists (and more recently automatically gathered wildsound) provided the scientists with a proving ground to evaluate different data processing systems to see how best to pull out and recognise an individual bird-song from amongst many sounds.
Analysis was made of recordings made in Brazil containing the sounds of over 500 birds, some of which may have not been individually recorded, so it was not an option to have a recognition system that compared recordings to a library of known birdsong. Rather the approach used 'learning algorathims' that led to the software identifying different songs.
The classification system performed well in a public contest using a set of thousands of recordings with over 500 bird species from Brazil. The system was regarded as the best-performing audio only classifier, and placed second overall out of entries from 10 research groups in the competition.
Dr Stowell commented,
There's a lot of interest at the moment in using automatic monitoring to measure how populations change over the years and from place to place. It won't replace the current methods of conservation monitoring but could add extra detail. For example, the British Trust for Ornithology uses a mixture of expert observations and crowd-sourced volunteer data to build up a picture of bird populations in the UK. Automatic monitoring could be a great additional source of information, filling out the picture for locations that haven't been manually checked.