Evaluating vowel normalization transformations through generalized linear modeling

Abstract

Nearey [ICSLP (1992)] suggests using generalized linear modeling (GLM) to investigate the role of vowel-intrinsic (e.g., role of F0) and vowel-extrinsic (e.g., z-transformation, log-mean transformation) normalization schemes. A generalized linear model consists of two parts: a linear predictor and a nonlinear link function. GLM is used to predict values both for dependent variables with discrete distributions and for variables that are nonlinearly related to the predictors. GLM can therefore be useful for dealing with relations between vowel-intrinsic and vowel-extrinsic information and vowel categories, since intrinsic factors can be tested as predictors in the linear predictor function and extrinsic factors can be tested using the link function. The GLM approach was applied to speech data from 80 speakers from four different regions in The Netherlands. The regional accent spoken in each group was different from the other three groups. The nine monophthong vowels of Dutch were produced in a neutral context. For each vowel, F0 and F1–F3 were measured. These values were transformed according to several vowel normalization schemes. The coefficients for the normalizations were estimated through maximum likelihood estimation. Overall, the results indicate that GLM can be a useful tool for evaluating normalization transformations.