Skip to Main Content
Various parameter sets-including a spectrum envelope, cepstrum, autocorrelation function, linear predictive coefficients, and partial autocorrelation coefficients (PAC's)- are evaluated experimentally to determine which constitutes the best parameter in spoken digit recognition. The principle of recognition is simple pattern matching in the parameter space with nonlinear adjustment of the time axis. The spectrum envelope and cepstrum attain the best recognition score of 100 percent for ten spoken digits of a single-male speaker. PAC's seem to be preferable because of their ease of extraction and theoretical orthogonalities; however, these PAC's tend to suffer from computation errors when computed by fixed-point arithmetic with a short accumulator length. We find two effective means to improve the errors; one is variable use of the PAC dimensions controlled by computation accuracy, and the other is smoothing along the time axis. With these improvements the PAC's offer almost 100 percent recognition.