Improvements to Percussive Component Extraction Using Non-Negative Matrix Factorization and Support Vector Machines


A system for the automatic extraction of percussive components from polyphonic digital audio is presented. Like some previous work, the system uses an iterative non-negative matrix factorization (NMF) algorithm to decompose a son spectrogram into components, and then it classifies these components as percussive or non-percussive using a support vector machine (SVM). Our approach attempts to reduce computation time and improve separation results by incorporating a perceptual dimensionality reduction into the NMF step. In addition, we introduce new featuresome based on note onset locationscted from the spectra and gain signals of each component in order to reduce classification errors. Our NMF approach greatly reduces computation time while retaining the same (or improving) the quality of separation. And using our new features, our component classifier achieves an equal error rate of less than 3.7% on a database of 32 songs.

Masters Thesis, UC Berkeley