Frequency warping for nonuniform talker normalization

This paper concerns a new approach to nonlinear spectral normalization to eliminate inter- speaker differences from frequency-band-limited speech. A frequency normalized distance between a test and a reference spectrum is defined on the basis of minimum mean square difference over all possible choices of frequency warping functions under certain constraints. This spectral distance is computed by means of dynamic programming after adaptively eliminating the individual glottal characteristics. Applications to the identification of steady-state vowels and the detection of words in connected speech is discussed.