Author(s): Matthew Aylett
E-mail: matthewa@ling.ed.ac.uk
In much experimental phonetics the concept of a vowel space is discussed. The density of points within this vowel space can be modelled using a mixture gaussian density function. There are a number of advantages in modelling the vowel space in this way. The model can generalise, producing a smooth representation of sometimes noisy data. The model can be used to normalise vowel spaces produced by different speakers. It can be used to categorise different vowels. It can be used to categorise and compare different types of speech, such as citation speech against running speech. To further improve the model a parametric curve fitting technique is applied. This paper describes the technique in detail, reports some tentative results and finally discusses the use of the model as a normalisation technique and as a method for measuring care of articulation.
Paper: Postscript (1033295 bytes)