Software performance for the automated identification of bird vocalisations: the case of two closely related species

Autonomous recording units now facilitate the large collection of audio recordings. However, the analysis of large amounts of acoustic data remains a challenge. The time required for manually searching for bird vocalisations may be equivalent or greater to the duration of audio recordings. This major constraint can be significantly reduced through the use of software developed for automated identification of bird vocalisations in audio recordings. We have compared the performance of four software (CallSeeker, Kaleidoscope Pro, Raven Pro, and Song Scope) and a Convolutional Neural Network (CNN) using audio recordings containing calls of Bicknell's Thrush and Gray-Cheeked Thrush, as well as the vocalisations of other bird species whose acoustic characteristics overlap with those of our target species. We evaluated all the software on the basis of two main criteria, their ability to detect calls and their ability to classify them correctly by species. Software performance ranged from 30 to 90% in terms of call detection (recall) and from 27 to 99% in terms of correct call classification (precision). CNNs offer a promising solution to the long-standing problem of detecting animal vocalisations in noisy soundscapes, while eliminating the tedious manual step of configuring the algorithms to maximise software performance.