Acoustic and perceptual speaker normalization

Abstract

An attempt was made to evaluate the performance of several methods for speaker normalization in both the acoustic and the perceptual domain of speech. This was done by comparing the acoustic distributions applied to the same vowel data and further by comparing the acoustic distributions to the perceptual distributions of the vowel data. The normalization methods included, among others, extrinsic (e.g., z-score transformation) and intrinsic methods (Bark transformation), and formant weighting (F2) and correction (F3−F2). To obtain the acoustic distributions, the normalization methods were applied to F0 and formant data from monophthong vowels in /sVs/ context of male and female speakers of Standard Dutch. The perceptual distributions were obtained through an experiment with phonetically trained listeners, whose task was to judge each vowel’s height, place of constriction and amount of rounding/spreading. The acoustic and perceptual distributions were compared using correlational- and cluster-analysis techniques. When describing the results of these tests, the focus will be on the extent to which the variation and overlap are the same in the acoustic and perceptual domains, and which normalization methods show a pattern of overlap and variation most similar in both domains. [Work supported by The Netherlands Organization for Research (NWO).]