In this paper we present a novel framework for real-time speech/music discrimination (SMD). The proposed method improves the overall accuracy of automatically classifying the signals into speech, singing, or instrumental categories. In our work, first, we design several groups of classifiers such that each group’s classification decision is biased towards a certain class of sounds; the bias is induced by training different groups of classifiers on perceptual features extracted at different temporal resolutions. Then, we build our system using an ensemble of these biased classifiers organized in a parallel classification fashion. Last, these ensembles are combined with a weighting scheme, which can be tuned in either forward-weighting or inverse-weighting modes, to provide accurate results in real-time. We show, through extensive experimental evaluations, that the proposed ensemble of biased classifiers framework yields superior performance compared to the baseline approach.
(See document for exact affiliation information.)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
