Central auditory model for spectral processing

Two central spectral processing models of the central auditory system for further processing the spatiotemporal firing pattern of the auditory nerve as outputs of the periphery auditory model are proposed. The temporal processing model capitalizes on the dominant periodicity in the firing pattern. The spatial processing model is based on the winner-take-all neural network mechanism and detects edges of discontinuities in the firing patterns. Experiments of noisy speech analysis showed that the proposed central auditory model can preserve prominent peaks at the formant resonances, while significantly reducing the output associated with the background noise. Noisy speech recognition experiments for both a small-scale classification task and a large-scale continuous speech recognition task showed that the proposed model provided a robust speech representation, and outperformed all other representations.<>