Singing Voice Detection using Modulation Frequency Features
Maria Markaki, André Holzapfel, Yannis Stylianou · 2008
In this paper, a feature set derived from modulation spectra is applied to the task of detecting singing voice in historical and recent recordings of Greek Rembetiko. A generalization of SVD to tensors, Higher Order SVD (HOSVD), is applied to reduce the dimensions of the feature vectors. Projection onto the “significant ” principal axes of the acoustic and modulation frequency subspaces, results in a compact feature set, which is evaluated using an SVM classifier on a set of hand labeled mu-sical mixtures. Fusion of the proposed features with MFCCs and delta coefficients reduces the optimal detection cost from 11.11 % to 9.01%. Index Terms: audio classification, modulation spectrum, singing voice activity detection.