On the Analysis of Predictive Data such as Speech by a Class of Single Layer Connectionist Models

Publisher Summary The chapter concentrates on the analysis of speech by connectionist models directly in the time domain, rather than in the frequency domain and includes the constraint that the input speech can be modeled by a linear predictive process. The chapter focuses on a class of networks where the order p of the linear predictive process is specified and produced an analytical solution for the weights of this class to minimize the network cost function. It is discussed that when the estimation errors are uncorrelated this solution is the same as that of conventional lp analysis. The few preliminary results explain that the nature of speech, with its correlated excitation mitigates against the weight values of the class corresponding exactly to the conventional lp coefficients, as predicted by the theory. However as in the case of conventional lp analysis, good classification of sounds has been indicated by preliminary results and this in turn has suggested a form of connectionist vector quantization (CVQ) structure. Finally the weight values of the class exhibit an interesting uniformity, which speculatively might have a physiological analogue.