Elliot Kermit Canfield-Dafilou
When a monophonic source signal is projected from two or more loudspeakers, listeners typically perceive a single, phantom source, positioned according to the relative signal amplitudes and speaker locations. While this property is the basis of modern panning algorithms, it is often desirable to control the perceived spatial extent of the phantom source, or to project multiple, separately perceived copies of the signal. So that the human auditory system does not process the loudspeaker outputs as a single coherent source, these effects are commonly achieved by generating a set of mutually decorrelated (e.g., statistically independent) versions of the source signal, which are then panned to make an extended source or multiple, independent source copies.
In this paper, we introduce an approach to decorrelation using randomly generated allpass filters, and introduce numerical methods for evaluating the perceptual effectiveness of decorrelation algorithms. By using allpass filters, the signal magnitude is preserved, and the decorrelated copies and original signal will be perceptually very similar. By randomly selecting the magnitude and frequency of the poles of each allpass biquad section in the decorrelating filter, multiple decorrelating filters may be generated that maintain a degree of statistical independence. We present results comparing our approach (including methods for choosing the number of biquad sections and designing the statistics of the pole locations) to several established decorrelation methods discussed in the literature.