Vowel‐space normalization procedures and phone‐preserving transformations of synthetic vowels

Abstract

A variety of speaker‐normalization procedures have been proposed for the recognition of vowels of different speakers and/or the cross‐language comparison of vowel systems. Many such procedures are special cases of affine transformations of coordinates in a two‐(or more) formant space. Results of perceptual experiments are presented indicating that only a limited subset of affine transformations preserve the phonetic identity of a set of synthetic vowels. From a perceptual point of view, many vowel normalization procedures appear to be too powerful in the sense that they can “over‐normalize”" vowels which are phonetically distinct. In particular, rotations of a two‐formant space and/or independent multiplicative scaling of the F1 and F2 axes lead to marked change in phonetic quality. On the other hand, results indicate that uniform multiplicative sealing of F1 and F2 by a single constant is “almost phone‐preserving” in the range of male/female vowel differences. For larger (e.g., male/child) differences, a specific nonuniform scaling similar to that proposed by Fant seems to be required to maintain “perceptual constancy.”