Scattering Transform of Averaged Data Augmentation for Ensemble Random Subspace Discriminant Classifiers in Audio Recognition

The paper presents an audio-based context recognition system using ensemble classifiers with wavelet feature extraction. The device-wise classification accuracy can be achieved with minimum computational time and improvement in classification accuracy compared to convolutional neural network (CNN) based technique via a subspace discriminant classifier. The proposed framework involves wavelet scattering feature extraction from the TAU2020 Mobile dataset via a modified data augmentation that improves the classification of unseen recording devices during training. The random sub-space discriminant classifier's device-wise classification performance on the extracted features is compared with other ensembles such as the bagged trees ensemble, boosted trees ensemble, and multiclass Naive Bayes classifier, and k-nearest neighbors. The evaluation results show around 72.6% accuracy and averaged 14.4% better than the ensemble methods. Compared to the highest performance of 76.5% in the DCASE2020 Challenge (Snapshot ensemble deep neural networks), the proposed approach offers a shorter computational time than CNN.