Spoken Digit Recognition Using Partial Autocorrelation Coefficients (PAC)

Abstract

Spoken digit recognition using autocorrelation coefficients (PAC or k parameters) is investigated with a digital computer. Speech samples are 10 spoken Japanese digits uttered by a single male speaker. Each digit is uttered 10 times. Each sample is analyzed at every 10 msec to give time series of PAC. Principle of recognition is a simple pattern matching with nonlinear time adjustment by DP procedure. When the dimension of PAC is greater than 4, recognition score is fairly good, but does not exceed the value of 98%. The cause of error seems to be the accuracy deterioration of computation of PAC. The computation of PAC is strongly correlated to a signal amplitude. Weighting PAC by amplitude information, the recognition score is improved and reaches at 100%, with the dimension of PAC greater than 4. When PAC is computed by fixed-point arithmetic with short accumulator length, the score become worse with the increase of the dimension. Further improvement of recognition score and time is expected by the variable use of PAC dimension controlled by the computation accuracy.