A series of upmixing techniques for generating multichannel audio from stereo recordings are proposed. The techniques use a common analysis framework based on a comparison between the short-time Fourier transforms of the left and right stereo signals. An interchannel coherence measure is used to identify time-frequency regions consisting mostly of ambience components, which can then be weighted via a nonlinear mapping function, and extracted to synthesize ambience signals. A similarity measure is used to identify the panning coefficients of the various sources in the mix in the time-frequency plane, and different heuristic mapping functions are applied to unmix (extract) one or more sources, and perceptually based functions to repan the signals into an arbitrary number of channels. We illustrate the application of the various techniques in the design of a two-to-five channel upmix system.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.