Spectral normalization of vowels based on frequency warping
Ghen Ohyama, Hisashi Wakita, Hiroshi Matsumoto · The Journal of the Acoustical Society of America · 1980
In a previous paper [J. Acoust. Soc. Am. Suppl. 1 64, S180 (1978)] speaker normalization based on a frequency warped spectral matching was discussed. A frequency normalized distance between test and reference spectra is defined on the basis of the minimum mean square difference over all possible choices of frequency. The slopes of the warping functions are restricted by an assumed range of the vocal tract length, so that they mostly take account of linear scaling for the difference in vocal tract length and of slight nonlinear difference. In this paper a modification of the technique will be proposed. Test and reference spectra are converted to a log scale frequency axis so as to give a heavy weighting in the region of the low frequencies and light weighting in the region of high frequencies. Then the distance between test and reference spectra is computed by means of dynamic programming, after eliminating the individual clottat characteristics. The experimental results show a better recognition score than previous ones.