Robust acoustic event classification using bag-of-visual-words

This paper presents a novel Bag-of-Visual-Words (BoVW) approach, to represent the grayscale spectrograms of acoustic events. Such, BoVW representations are referred as histograms of visual features, used for Acoustic Event Classification (AEC). Further, Chi-square distance between histograms of visual features evaluated, which generates kernel to Support Vector Machines (Chi-square SVM) classifier. Evaluation of the proposed histograms of visual features together with Chi-square SVM classifier is conducted on different categories of acoustic events from UPC-TALP corpora in clean and different noise conditions. Results show that proposed approach is more robust to noise and achieves improved recognition accuracy compared to other methods.