A Modified Spatio-Temporal Orthogonal Iteration Method for Multichannel Audio Signal Representation

In this paper, we present a novel algorithm for a spatio-temporal extension of the well-known method of orthogonal iterations in linear algebra. This algorithm estimates an n-input, m-output (m ≪ n) paraunitary filter bank from a multichannel data auto-correlation sequence to maximize the total output power of the filter bank when applied to an n-dimensional input signal. We then show how this procedure can be used to generate reduced rank signal representations of recordings of m audio sources in a room as collected by an n-channel microphone array. The importance of the method for determining the number of active sound sources in a room for convolutive blind source separation is also discussed.