Machine Learning-Based Sound Event Detection: A Case Study for Noise Identification in Classroom Environment

Peaceful academic surroundings are one the most fundamental requirements in any teaching learning environment. The main aim of this paper is to explore the applicability of sound event classification for noise in a classroom environment. Sound event classification in a multi-source environment becomes more challenging due to the presence of polyphonic and overlapping sound. For a complex environment, results of domain-specific sound event detection are not much accurate as it is difficult to fully extract features from models with a single input. In this paper, deep and acoustic feature-based classification of sound events is performed using a CNN-based model and a transfer learning approach using YAMNet. In this study, we introduce a new dataset which we call the school classroom dataset (SCD). Also, in addition to using SCD, we have performed the sound event classification on the standard Urban Sound 8K Dataset and ESC-50 datasets. The experimental results demonstrate that the melspectrogram features along with MFCC and Chroma features are better suited for improving the model performance and achieve better performance over the other models in the literature.