Experimental Upper Bound for Convolutive Mixtures of Speech by Maximizing SIR

A new experimental upper bound to evaluate the separation performance of blind source separation methods for (convolutive) mixtures of audio is proposed. The upper bound produces an estimate of the best separation performance for a specified linear feedforward structure and it gives the set of parameters required to achieve the optimal performance. It also can be used to measure the absolute, as opposed to relative, separation performance for a given blind source separation algorithm. The upper bound is obtained by maximizing directly the (mean) signal-to-interference ratio (SIR). The new method is an improvement over a previously proposed experimental upper bound, obtained by maximizing an approximation to the SIR. Experimental evaluation is performed which indicates the usefulness of the method.