ABSTRACT
This chapter provides a brief introduction to typical neural network structures and their application to speech recognition tasks. The axons pass in proximity to the input dendrites of other neurons, and the coupling from a pulse in the axon from one neuron will stimulate the tendency to fire in the others. A very obvious way in which operation of the brain differs from the operation of normal computers is the degree of parallel processing. The nature and properties of the brain have inspired many research groups to investigate whether cognitive processes could be achieved in electronic or computational models that have many of the known neural properties. The motivation for the development of hybrid models for continuous speech recognition is to combine the discriminative classification abilities of Artificial neural networks with the time-domain modelling capabilities of hidden Markov models (HMMs). The usual approach involves training a neural network to compute emission probabilities in an HMM system.
