Hearing aids for the profoundly deaf based on neural net speech processing

A new speech processing concept for cochlear implant (CI)-systems has been developed. It is based on robust feature extraction and a neural net classifier. Feature coefficients, extracted either by relative spectral perceptual linear predictive technique or regular CI-filtering, are classified into `auditory related units´. The classifier is based on an adapted self-organizing Kohonen (1990) algorithm which finds representative clusters in the input feature vector space. These clusters are closely related to the statistical distribution of the feature coefficients and represent phonetic units. Firing neural net output nodes control the synthesis of a limited `stimulus pattern alphabet´. Each `letter´ represents a sub-phoneme and is linked to a highly distinguishable complex stimulus pattern. The concept has been implemented with CINSTIM V2.0. First experimental results confirm the new CI speech processing strategy