Neural network based speech recognition method utilizing spectrum- dependent and time-dependent coefficients

摘要

An improved artificial neural network for use in speech recognition is disclosed. It comprises an input layer, a hidden layer, and an output layer, each of these layers consisting of a plurality of nodal points. A set of first weighting coefficients are used between the input layer and the hidden layer which are functions of at least one of the nodal points in the hidden layer and at least one of the nodal points in the input layer; whereas, a set of second weighting coefficients, which are functions of time and at least one of the nodal points in the output, are used to correlate between the hidden layer and output layer. In a preferred embodiment, the first weighting coefficients are calculated using the following formula: ##EQU1## i is the index for nodal point in the input layer and a.sub.j, b.sub.j, and c. sub.j are all training coefficients associated with nodal pointj in the hidden layer; and the second weighting coefficients are calculated using the following formula: ##EQU2## n is the timeframe number, r is the order of an orthogonal polynomial series (&psgr;, 60 .sub.jkm is the m-th order training coefficient between nodal points j and k, in the hidden and output layers, respectively. The use of the two different sets of weighting coefficients allows a timeframe-based division of the speech signals, resulting in a substantial reduction of parameters required for accurate speech recognition.