Sound event classification using neural networks and feature selection based methods

Abstract

Sound events, emanate from several sources, are ubiquitous and manifest them selves with different characteristics in different environments. With the advancement of deep learning models and existence of ever increasing training data, the automatic recognition and classification task of these events has improved significantly over the years . Traditionally, environmental sound event recognition systems are developed by keeping the generic database that is readily available, while sound events generated in a particular environment are not focused. Another issue of training of large neural networks requires huge amount of parameters and training them costs computational resources. To tackle this issue, we firstly built a custom database consisting of events occurring outside and around smart homes or building. The sound events such as rain, wind, human gait, and passing of vehicles. We propose the use of a sequential feature selection technique for for reduction of dimension of features extracted with MFCC. Selected features are used for training recurrent neural network (RNN) on aforementioned sound events. We compared the results of our proposed method with the same RNN trained with MFCC features and convolutional neural networks (CNN) trained with mel frequency band (MFB) features. Our proposed system performed with high accuracy in former case but slightly better compared to CNN in achieving higher classification accuracy and a significant reduction of parameters during training with the proposed system.