Pulse-induced masking of sonagrams for the enhancement of speech

Speech-enhancement on the basis of sonagrams and an artificial neural network is investigated. We utilise a biologically inspired model - a Pulse Coupled Neural Network - and its temporal segmentation property. Pulses of linked neurons and their temporal as well as their spacial relation regarding the receptive field induce masks, providing characteristic elements for a reduced signal-representation. Properties of the model and its integration into an enhancement process are presented. Experiments on speech and non-speech signals are evaluated and discussed with regard to observed signal processing capabilities.