Simple Speech Recognition with Little Linguistic Creatures

In this paper we present a stochastic neural­network architecture, the synchronous-network acceptor, that can approximately simulate nonde­ terministic finite-state automata, where the preci­sion of the approximation depends on the level of noise in the synchronous-network acceptor. This network learns to simulate finite-state automata by means of an unsupervised learning algorithm. The synchronous-network acceptor is a neuro­physiologically plausible model of connected relay nuclei, or connected, coherent, collections of cor­tical columns. The learning algorithm is a plausi­ble model of a column formation process or, more generally, an axon terminal segregation process. Complex self-organizing systems can be con­structed as ensembles of synchronous-network ac­ceptors. In virtue of the fact that finite-state automata can be simulated, applications of such systems can be implementations of linguistic the­ories, for instance parsing algorithms. However, we prefer a reductionistic, holistic approach. The approach is reductionist in that linguistic behav­ior is studied via its neurophysiological substrate. Our approach is holistic because we consider com­plete linguistic systems. These systems are called little linguistic creatures. They typically contain models of nervous sensory, motivational, and mo­tor subsystems, and unlike in more conventional approaches in which various linguistic faculties are studied separately, we reduce the complexity of the subject by modeling simple linguistic sys­tems. This holistic approach should in the end have the advantage of domain independence over more conventional approaches. The system of self-organizing synchronous network acceptors we present in this paper is a very simple linguistic creature for speech recog­nition. Its state provides the semantics of the linguistic input and the context information for the resolution of nondeterministic choices