The application of artificial neural networks to provide safe access to driver information systems and other noncritical automotive functions

The need for intrinsically safe access to driver information systems and other noncritical automotive functions is a pressing one and needs no elaboration. Voice access systems have been proposed for this. The authors describe the results of an initial investigation into the use of time encoded speech (TES) fixed size, fixed dimension A-matrices as the input layer of a simple neural network, configured to recognise the digits `zero´ to `nine´ from a male and from a female speaker. Input speech data recorded in an environment of 75 dB SPL background noise is first formed into `A´ matrices and then presented to the network. Subsequently, this same data is partially mutilated to mimic, to a first order, the effects of additional acoustic background noise. The network involved is a modification of the back error propagation algorithm based on the generalised delta rule