Frequency-axis warping to improve automatic word recognition

Frequency normalization of talkers remains a problem in word recognition, especially where new talkers cannot be asked to provide samples (of their vowels, for example) in advance. Several methods were investigated; for each, parameters were derived by calculating their effect on formant histograms derived from casual speech. Methods tried were a) uniform multiplication of frequencies ("stretching" the vocal tract); b) "stretching" each formant region by a different amount; c) combined shift and stretch (affine mapping); d) different affine mappings for different formants (this includes warping each formant as a function of its range); e) warping each formant non-linearly as a function of its distribution. Experiments show that parameters derived from casual speech improve vowel recognition markedly, and that method e) appears strongest.