A Deep Meta-model for Environmental Sound Recognition

Nowadays, sound serves as a crucial factor in all facets of human life. Staring from automating personal security systems to critical surveillance systems, sound is an indispensable component. The practical implementation of the present day automatic sound recognition systems in real-life settings is inadmissible due to their poor detection accuracy. However, deep learning-based systems overcome the incompetence of the traditional machine learning-based models, and it can be used to develop automatic sound classification systems. This work proposes a deep meta-model for categorizing environmental sounds on the basis of the spectrogram images generated from these sounds. In the proposed approach, spectrogram images of environmental sounds are used to train five different deep learning models, and the predictions from these base models are then stacked using the proposed deep meta-model. Experimental results on two benchmark datasets such as ESC-50 and UrbanSound 8K demonstrate the fact that the proposed deep meta-model is a promising alternative to the conventional approaches for environmental sound recognition.