Application of the dempster-shafer theory of evidence to improved-accuracy isolated-word recognition

This paper describes experiments in which the outputs of three isolated-word recognition algorithms were combined to yield a lower average error rate than that achieved by any individual algorithm. The input tokens were simulated, fixed-length spectral speech patterns subjected to additive noise and either a "high" or "low" degree of time distortion. The three recognition techniques used were dynamic time warping, hidden Markov modelling and a multilayer perceptron. Two combination techniques were employed: the formula derived from the DempsterShafer (D-S) theory of evidence and simple majority voting (MV). D-S performed significantly better than MV under both time-distortion conditions. Evidence is also presented that the assumption of independent word scores, which is necessary for D-S theory to be strictly applicable, is questionable.