Download original image
Structure of the proposed externalization model, consisting of a short-term and a long-term memory. In the long-term memory, the spectral gradients and ILDs are extracted from the direct sound part of the target signal in each frequency channel of a gammatone filter bank. In the short-term memory, the ILD temporal standard deviations are obtained from the echo-suppressed reverberant signals in each frequency channel. The deviations of these three acoustic cues from the template signals are summed up with different weighting factors and mapped to externalization ratings.