This study compares the ability of four classification methods to distinguish between songs of individual Mexican Antthrush Formicarius moniliger: self-organizing maps (SOMs), discriminant function analysis, fuzzy logic and hidden Markov models. Recordings were made under field conditions in a Mexican rainforest. Two types of data were analysed - recordings from birds that had been ringed and identified to sex, and recordings from birds that had been identified based on their recording location and song timing. An event detector extracted song features and SOMs were used to confirm the number of individuals recorded. The SOM separated all five ringed birds successfully, and also differentiated two other birds that were not identified while vocalising. The three supervised learning methods correctly classified over 97% of songs to individual from the set of identified recordings. Tests with songs for predicted, rather than known, individuals yielded more variable results across methods, with results ranging from 77.8% to 93.9% correctly identified. The respective merits of the three supervised classification procedures are discussed for automated recording, detection and classification.
bioacoustics, Formicarius, self-organizing maps, hidden Markov models, vocal individuality