Unsupervised neural networks for speech perception with Cochlear Implant systems for the profoundly deaf

Abstract

Recently we have proposed a new speech processing concept for Cochlear Implant (CI) — systems. The concept is based on speaker independent signal representation and a neural net classifier which can be combined with the well known CI- speech-coding-strategies. This paper describes some new simulation results: For every speech input frame a 4- dimensional feature vector has been extracted by employing a relative spectral perceptual linear predictive (RASTA-PLP) technique. To classify the feature vectors into so called “auditory related units (ARU)” we applied the self-organizing Kohonen neural net The best matching ARU's will directly control the synthesis of a “alphabet” of patient adapted stimulus patterns. Simulation results show that the Kohonen algorithm finds representative clusters in the feature vector space for different net dimensions. A discussion of the results and a overview of present experiments with deaf patients will be given.