Fully and semi-automated analyses of bird song based on machine learning approaches [abstract]

The estimation of birds’ repertoire size is based on spectrographic analysis. This is the most important feature and used to describe song complexity and organization. The estimation of repertoire size is time consuming and involves the objective scoring of song elements by eye. Moreover, based on such manual approaches, it is difficult to make quantitative comparisons between individuals, because it necessitates a universal coding system of song elements at the population level, for which the observer would need to handle cognitively all the song elements found in the population. To enhance this process, we have developed a computerised scoring system, and created a sound library of song elements that is universally applicable to all 67 individuals in male collared flycatcher (Ficedula albicollis) recorded in Hungary. The manual estimation of repertoire size followed the visual scoring of spectrograms, while we wrote a computer program in Matlab environment for the automatic estimations. Based on some acoustic features (time and frequency parameters) measured automatically, the program offers a semi- and a fully automatic way to insert each syllable into the universal syllable library. The semi automatic approach relies on a decision tree method, which shows the most appropriate matches between syllables, but always request a decision from the investigator to assign each syllable. In the fully automatic way, the program uses a self-organizing map method and performs the syllable scoring directly. After the comparison we found that the automated approaches can provide reliable and repeatable estimates on song variables. The automated methods can be effectively exploited in future studies of song complexity and organisation in birds, or even other animals.