Sound Event Detection Research Based On NMF

In this paper, a sound event detection system that can recognize and temporally locateoverlapping sound events in recordings was proposed. In general, studies related to soundevent detection with overlapping events that are explicitly annotated, but the detection resultsare presented as a sequence that is assumed to contain only the most prominent event at eachtime. However, some important information is contained in the other sound events. In thispaper, the overlapping sound events detection based on the isolated sound events detectionwas studied, but no only detect overlapping sound events by the isolated sound eventsdetection algorithm. A solution was proposed that combines NMF and GMM to detectoverlapping sound events. Firstly, the signals are preprocessed using an unsupervisednon-negative matrix factorization based algorithm in order to separate sound sources intotracks. Secondly, event detection based Gussian Mixture Model is performed on each track.At last, this solution was tested with two databases, namely an isolated sound events databaseand an overlapping sound events database. The results show that GMM performs better thanHMM or only NMF in isolated events detection problem. However, when overlapping eventsare detected by GMM, the performance is falling down. The performance is so good in detectoverlapping sound events based the solution proposed in this paper. At the same time, thealgorithm proposed in this paper is better than other similar papers.