Classification of time-frequency regions in stereo audio

This paper is about classification of time-frequency (TF) regions in stereo audio data by the type of mixture the region represents. The detection of the type of mixing is necessary for example in source separation, upmixing and audio manipulation applications. We propose a generic signal model and a method to classify the TF regions into six classes which are different combinations of central, panned, and uncorrelated sources. We give an overview of traditional techniques for comparing frequency-domain data and propose a new approach for classification which is based on measures specially trained for the six classes. The performance of the new measures is studied and demonstrated using synthetic and real audio data.