Active learning for classifying long‐duration audio recordings of the environment

Abstract

Abstract This paper presents an active learning (AL) framework for the classification of 1‐min audio recordings derived from long‐duration recordings of the environment. The goal of the framework was to investigate the efficacy of AL on reducing the manual annotation effort required to label a large volume of acoustic data according to its dominant sound source, while ensuring the high quality of automatically labelled data. We present a comprehensive empirical comparison through extensive simulation experiments of a range of AL approaches against a Random Sampling baseline for soundscape classification. Random Forest is used as a benchmark supervised approach to build classifiers in the AL framework. Also, 12 summary indices extracted for each 1 min of 13‐month recording are used as features for training the classifiers. Our experimental findings demonstrate that (a) among existing query strategies, those based on classifier confidence and diversity of samples are more effective for very large datasets where the classes are imbalanced in size; (b) by considering a practical target performance (i.e. F‐measure equal or greater than 0.8, 0.85 and 0.9) for AL, only 5–16 hr of manual annotation effort is required to build a classifier that automatically annotates a large amount (13 months) of unlabelled audio data. Active learning has a key role to play in alleviating the burden of manual annotation required to build classifiers which can support effective monitoring of species diversity in at‐risk ecosystems.