NMF-based temporal feature integration for acoustic event classification

In this paper, we propose a new front-end for Acoustic Event Classification tasks (AEC) based on the combination of the temporal feature integration technique called Filter Bank Coefficients (FC) and Non-Negative Matrix Factorization (NMF). FC aims to capture the dynamic structure in the short-term features by means of the summarization of the periodogram of each short-term feature dimension in several frequency bands using a predefined filter bank. As the commonly used filter bank has been devised for other tasks (such as music genre classification), it can be suboptimal for AEC. In order to overcome this drawback, we propose an unsupervised method based on NMF for learning the filters which collect the most relevant temporal information in the short-time features for AEC. The experiments show that the features obtained with this method achieve significant improvements in the classification performance of a Support Vector Machine (SVM) based AEC system in comparison with the baseline FC features.