ICA-based Single Channel Audio Separation: New Bases and Measures of Distance

Independent Component Analysis (ICA) can be used for single channel audio
separation, if a mixed signal is transformed into time-frequency domain and the
resulting matrix of magnitude coefficients is processed by ICA. Previous works used
only frequency (spectral) vectors and Kullback-Leibler distance measure for this
task. New decomposition bases are proposed: time vectors and time-frequency components.
The applicability of several different measures of distance of components
are analysed. An algorithm for clustering of components is presented. It was tested
on mixes of two and three sounds. The perceptual quality of separation obtained
with the measures of distance proposed was evaluated by listening tests, indicating
“beta” and “correlation” measures as the most appropriate. The “Euclidean” distance
is shown to be appropriate for sounds with varying amplitudes. The perceptual effect
of the amount of variance used was also evaluated.