Classification of Environmental Sounds with Convolutional Neural Networks

The use of sound data is critical for predicting the effects of environmental activities and gathering information about the environment of these activities. Sound data is utilized to obtain basic information about the functioning of urban activities such as noise pollution, security systems, health care, and local services. In this sense, Environmental Sound Classification (ESC) is becoming critical. Due to the increasing amount of data and time constraints in analysis, there is a need for new and powerful artificial intelligence methods that enable instant automatic identification of sounds. These methods can be developed with Convolutional Neural Networks (CNN) models, which have achieved high accuracy rates in other fields. For this reason, in this study, a new CNN based method is proposed for the classification of two different CSR datasets. In this method, the sounds are first converted into image format. Then, novel ESA models are designed for the classification of these sounds in image format. For each dataset, the ESA models with the highest accuracy rate were obtained among the multiple ESA models designed. The datasets used in the study are ESC10 and UrbanSound8K, respectively. The sound recordings in these datasets were converted to image format with 32x32x3 and 224x224x3 dimensions, and four different image format datasets were obtained. The CNN models developed to classify these datasets are named ESC10_ESA32, ESC10_ESA224, URBANSOUND8K_ESA32, and URBANSOUND8K_ESA224, respectively. These models were trained on the datasets using 10-fold cross-validation. In the obtained results, the average accuracy rates of the ESC10_ESA32, ESC10_ESA224, URBANSOUND8K_ESA32, and URBANSOUND8K_ESA224 models are 80.75%, 82.25%, 88.60%, and 84.33%, respectively. When these results are compared with other baseline studies in the literature on the same datasets, it is seen that these models achieve better results.