Coding of speech into nerve-action potentials

Abstract:
One of the most critical processing steps during encoding of sound signals for neuronal processing is when the analog pressure wave is coded into discrete nerve-action potentials. This conversion induces massive information loss - or to phrase it positively - information reduction. As any information lost during this process is no longer available for neuronal processing, it is important to understand and quantitatively model the underlying principles. We have developed a detailed model of auditory processing, which codes sound signals into spike-trains of the auditory nerve. These drive Hodgkin-Huxley models of cochlear nucleus neurons, which are known to extract temporal features of sound signals.

We analyze the quality of coding with the framework of automatic speech recognition and the temporal information processing capabilities with the methods of information theory. Our latest improvements in speech coding by introducing the effect of offset-adaptation together with an improved matching of neuronal features to the speech recognizer using an artificial neuronal network has lead to significant improvements of recognition scores, now reaching the values of successful technical feature extraction methods. Offset adaptation is also required to drive onset neurons in the cochlear nucleus, which are able to code temporal information with sub-millisecond precision (< 0.02 ms). Our results provide quantitative insight into temporal processing strategies of neuronal processing and are highly relevant for cochlear implants.

Werner Hemmert, Marcus Holmberg, Huan Wang, Michele Nicoletti, Michael Isik, Sonja Karg, Willy Bade Bio-Inspired Information Processing, IMETUM - Institute for Medical Engineering, Technische Universit�t M�nchen and Bernstein Center for Computational Neuroscience, Munich

back