Temporal alignment, spatial spread and the linear independence criterion for blind separation of voices

The usefulness of the neural network method of Matsuoka et al. [1995] for separating a mixture of two signals is investigated. The method appears to be very effective at separating signals which have been combined synthetically, but much less effective at separating a mixture of two real voices recorded with a pair of microphones. The algorithm was applied to specific examples to determine how critical it is that they be temporally aligned and that there be no spatial spread of the sources. The results indicate that the algorithm is very sensitive to temporal misalignment of voice mixture signals, whilst the spatial spread of the voice sources is less significant. This suggests that adaptive alignment of the mixture signals before signal separation may be beneficial