Environmental Sound Monaural Source Separation with Clustered Non-Negative Matrix Factorization

With the increase in single channel environmental noise monitoring data sources, the need for monaural blind source separation (BSS) has also increased. There is great potential in this acoustic data analysis area to learn about an environment by picking out the individual sources creating sound. We propose using non-negative matrix factorization (NMF) to extract simple signal parts and then combining them to resemble original sources using clustering methods. We examine the effects of three parts of the clustering process on performance: input type, dissimilarity measure, and clustering method. Each simple part extracted with NMF has a time and frequency component which can be used as an input to the dissimilarity calculated between all of the extracted signals. In this work we study the Earth Mover's Distance and Dynamic Time Warping for their lenient alignment properties as well as Euclidean Distance as a baseline comparison. For the clustering methods, we look at the popular k-medoids and hierarchical clustering along with Deterministic Annealing. In terms of performance, clustering was shown to improve results over the non-cluster NMF method and in terms of understanding, examining the performance of the several variants described above allowed us to draw conclusions on salient acoustical properties.