The Analysis for the New Individualized Features Derived from Finite Ridgelet Transform

In the literature of speaker recognition, the short-term features obviously dominates in the description of the individualized information for the candidates whose conventional assumption is the nonstationarity in the block. This paper discards the popular ideas and produces such features as segment center decimation(SCD), differential segment center decimation (DSCD), maximum element(MAE) and minimum element(MIE) to examine the geometrical composition of the spectrum in the time-frequency plane of the one-dimensional speech signal. A great number of the experiments based on these features have shown that the feature sets of segment center decimation and differential segment center decimation possess the favorable recognition performance respectively, especially when the process of the reasonable dimension reduction is imposed