Classroom Activity Detection in Noisy Preschool Environments with Audio Analysis

Pre-school teaching demands several diverse methodologies making the quantification of classroom activity very useful to reflect on and refine teaching practices. Therefore, automating the monitoring process with unobtrusive audio recording and analysis is viewed as a simple and scalable solution. We present a system for performing automatic classroom activity detection from audio using a new audio corpus obtained from preschool classrooms in rural India. We define the sound classes of interest in estimating the level of teacher engagement. To address the challenge of training data scarcity, we investigate the combination of pre-trained models with unsupervised homogeneity-based segmentation to realise a multilabel classification scheme. We discuss the performance of the classifier and its utility in the context of the classroom activity monitoring task.