Active learning for acoustic classification

Classification of noise sources based on their acoustic signatures is becoming increasingly necessary due to the increasing prevalence of noise monitoring systems. The intractable amount of data that these systems capture necessitates robust and efficient classification algorithms. However, the process of building a classifier requires manually labeling recorded noise events, which is a time consuming task. In this paper we explore a method for reducing the human burden and making the process more efficient called active learning, or “listener-in-the-loop.” This method iteratively improves classifier performance by querying a human listener using optimally chosen observations based on certainty factors, which measure the degree of belief or disbelief based on probability for a particular classification. Classifier performance as well as algorithm efficiency, in terms of computational costs and amount of data required, will be discussed.