Ensemble and Multi Model approach to Environmental Sound Classification

AI plays an important role in acoustics recognition. Importantly, being able to automatically and accurately identify environmental sounds opens up a broad range of applications. Deep learning techniques can assist in the recognition of sounds which we come across in our day-to-day life. Most of the previous work in environmental sound classification involves training a model on a single set of features. In our work, we extract two sets of features from the audio namely Mel spectrograms and Mel Frequency Cepstral Coefficient Spectrograms. We propose two CNN-based approaches through the usage of Ensemble of models which are trained on Mel Spectrograms and a multi model network which is trained on both Mel and MFCC Spectrograms. The proposed approaches were trained and tested on the ESC-50 dataset which contains 2000 samples of audio recordings distributed into 50 categories. Performance evaluation revealed that both our proposed approaches achieve an accuracy of 92%.