Max-Fusion of Random Ensemble Subspace Discriminant with Aggregation of MFCCs and High Scalogram Coefficients for Acoustics Classification

In this paper, a random sub-space discriminant classifier for classifying acoustic devices that combines the features obtained from Mel-frequency cepstral coefficients (MFCCs), and scalogram coefficients is proposed. The aggregated features for the random ensemble sub-space discriminant classifier model are used. The maximum weight fusion mechanisms are used to fuse the ensemble classifier’s results from the two sets of coefficients. With a higher concentration of scalogram coefficients, the accuracy improves by around 8.15% compared to single feature extraction via MFCCs. The random subspace discriminant classifier achieves the classification accuracy of approximately 74.9% or 20.8% better than the baseline result of 54.1% obtained in the DCASE2020-Task1A Challenge.