Acoustic classification of frog within-species and species-specific calls

There have been various studies using automated recognisers of acoustic features and machine learning algorithms to classify frog species within a chorusing community. Such studies rarely consider within-species call variation in the classification process. Individual frog species may make a range of different calls, with different purposes. Including modification of calls in automated recognition has the potential to not only increase the accuracy of classification of calls to species, but also to provide information on frog calling behaviour within species. Here we use acoustic feature extraction and machine learning algorithms (1) to investigate the acoustic feature importance of identifying species-specific calls; (2) to determine which acoustic features can be used to classify within-species calls. Our method was tested for its performance in recognising four frog species (Litoria bicolor, Litoria rothii, Litoria wotjulumensis, and Uperoleia inundata) and four call types of L.wotjulumensis (normal, click, response, and long trill). Mean classification accuracy was high, with 84.0% at the species level, and 83.7% at the call type level. The overall classification accuracy can be up to 93.0%, when considering four call types of L. wotjulumensis as individual classes and being combined with other three frog species. Two techniques, principal component analysis and Fisher discriminant ratio were used for dimension reduction and to select important features for discriminating among calls of different species, and call types within species. In conclusion, our proposed classification mechanism could effectively not only classify different frog species but also identify different call types within the same species. Moreover, we found that time-domain features were important for classification of within-species calls, whereas frequency-domain features were more useful for classification of species-specific calls.