Comparison of coarse and fine sampling resolutions in vowel analysis

Abstract

There are many researcher degrees of freedom in phonetic analyses, especially in that of segments such as vowels. The focus of this study is one such degree of freedom, governing how frequently measurements are taken for an individual token in the corpus. Choices for this stage of analysis include averaging formants (single measurement), taking point measurements (multiple measurements), or sampling continuously across the token (many measurements). Less-frequent measurements are convenient techniques compatible with many common statistical methods, but risk oversimplifying the patterns in the data. In this project, I compare point and continuous measurements using Generalized Additive Mixture Models (GAMMs) on speech corpus data. GAMMs are a generalized model that can flexibly parameterize both curves and points based on a combination of random and fixed effects. To compare different measurement frequencies, I quantify the benefit of additional measurements in terms of information gain and variance explained. I also address the modeling of multidimensional measurements and identify the degree to which issues of multidimensionality such as collinearity can be accounted for in GAMMs.