Evaluating vowel normalisation procedures: A case study on Southern Sotho vowels

Abstract

Extended research has established the fact that it is difficult, if not impossible, to conduct a direct acoustic–phonetic comparison between the vowels produced by female and male speakers or groups of speakers. This is due to the effects of physiological and anatomical speaker differences on vowel formants, especially F1 and F2. A variety of computational procedures (or algorithms, methods, techniques) have been developed to allow for the normalisation of formant-frequency measures and to assist in reducing or even eliminating the consequences of speaker-specific variability, without losing information with regards to other important features, such as age and ethnicity. In the present article, eight such procedures are tested on a large data-set of Southern Sotho vowels (4 434 tokens), as produced by twelve speakers (six of each gender and balanced as to age and locality). We concentrated on the examination of two types of normalisation procedures: firstly, the vowel-extrinsic class of normalisation procedures, which requires a sampling of the entire vowel space of the language; and, secondly, the vowel-intrinsic class, the relevant procedures of which can be conducted on a single vowel or vowel class. In this study, the various procedures are evaluated by using a squared coefficient of variance (SCV) metric and then converting the results to effect sizes, in terms of d-scores. The results in all instances show Lobanov's procedure (a vowel-extrinsic, formant-intrinsic procedure) to be superior to the others, and Bark (vowel-and formant-intrinsic) to be the worst performer. In the case of vowel-extrinsic procedures, the relative strength of two procedures, both developed by Nearey, was evaluated; one of them is formant-intrinsic, the other is formant-extrinsic. All things being equal, the former, as predicted on the grounds of evaluations reported in the existing literature, turned out to be the better performer.