Research on a novel Saliency map computational model of auditory attention

A key problem of the computational model of exogenous auditory attention is how to calculate the saliency map of acoustic signal, in order to detect the saliency of speech signals, an effective bottom-up auditory saliency map computational model is proposed in this paper. To get this model, the auditory nerve response of each frequency channel which processed by the peripheral auditory system was obtained firstly, and divided them into sub-frames to calculate short-term energy, then，basing on the characteristics of auditory stream， energy spectrum of each channel was time-domain filtered in different scales through Gaussian filter groups. Afterwards auditory saliency of each frequency channel was computed by using center-surround differences operator, and the auditory saliency map was achieved by linear combination of each frequency channel saliency finally. The effectiveness of the proposed model was validated via the simulation experiments of different acoustic signals and SNR. Comparing with the traditional model, the simulation results and analyses indicated that the proposed model could calculate the saliency map of acoustic signal more accurately and highlight the saliency of signal mutation moments, also had fine applicability to low SNR signals.