Machine Learning Algorithms for Automatic Classification of Marmoset Vocalizations

Abstract

Automatic classification of vocalization type could potentially become a useful tool for acoustic the monitoring of captive colonies of highly vocal primates. However, for classification to be useful in practice, a reliable algorithm that can be successfully trained on small datasets is necessary. In this work, we consider seven different classification algorithms with the goal of finding a robust classifier that can be successfully trained on small datasets. We found good classification performance (accuracy > 0.83 and F1-score > 0.84) using the Optimum Path Forest classifier. Dataset and algorithms are made publicly available.

Fig 1. Vocalization exemplars.

Amplitude and time-frequency spectrograms are shown for representative exemplars of the marmoset call types considered in this study. A: Alarm, B: Chirp, C: Loud shrill, D: Phee-2, E: Phee-3, F: Phee-4, G: Seep, H: Trill, I: Tsik, J: Tsik-Ek, K: Twitter.

Fig 2. The effect of training set size on classification performance.

For the sake of visual clarity, the results of OPF using the distance metrics Bray-Curtis and Chi-Square, and SVM using linear and polynomial kernels are excluded.