| Traditional methods of speech recognition have very limited complexity and impose considerable grammar constraints. Today's systems have critical problems understanding different voices and do not have robust vocabularies. This thesis describes research on a biological method for speech recognition that models an artificial cochlea using digital LTI filters. The mammalian cochlea has unique filtering characteristics that are highly non-linear with the level of input sound. In this research, we designed a non-linear cochlear filter which dynamically changes its response characteristics in approximately the same way the mammalian cochlea does. For large input levels, the filtered signal is relatively flat with cut off frequencies below the tuned frequency, while for lower input levels it is highly peaked at the tuned frequency. |
