Robust speech recognition with dynamic synapses

摘要

We have developed a speech recognition system employing the concept of dynamic synapses. A dynamic synapse incorporates fundamental features of biological neurons including presynaptic mechanisms influencing the probability of neurotransmitter release from an axon terminal. With these mechanisms, the probability of neurotransmitter release becomes a function of the temporal pattern of action potential occurrence, and hence, transforming a spike train into a sequence of discrete release events. When presynaptic mechanisms vary quantitatively across the axon terminals of a single neuron, an array of spatially distributed temporal patterns can be generated. In other words, information is coded in the spatio-temporal patterns of release events which provides an exponential growth of coding capacity for the output signals of a single neuron. A dynamic learning algorithm is developed in which alterations of the presynaptic mechanisms lead to different pattern transformation functions while changes in the postsynaptic mechanisms determines how the synaptic signals are to be combined. We demonstrate the computational capability of dynamic synapses by performing speech recognition from unprocessed, noisy raw waveforms of words spoken by multiple speakers with a simple neural network consisting of a small number of neurons connected with dynamic synapses. The system is highly robust against noise, and outperformed human listeners under some conditions.