Perceptual validation of vowel normalization methods for variationist research

Santiago Barreda

doi:10.1017/s0954394521000016

Santiago Barreda

Open Access

https://doi.org/10.1017/s0954394521000016

Copy DOI

Abstract

AbstractThe evaluation of normalization methods sometimes focuses on the maximization of vowel-space similarity. This focus can lead to the adoption of methods that erase legitimate phonetic variation from our data, that is, overnormalization. First, a production corpus is presented that highlights three types of variation in formant patterns: uniform scaling, nonuniform scaling, and centralization. Then the results of two perceptual experiments are presented, both suggesting that listeners tend to ignore variation according to uniform scaling, while associating nonuniform scaling and centralization with phonetic differences. Overall, results suggest that normalization methods that remove variation not according to uniform scaling can remove legitimate phonetic variation from vowel formant data. As a result, although these methods can provide more similar vowel spaces, they do so by erasing phonetic variation from vowel data that may be socially and linguistically meaningful, including a potential male-female difference in the low vowels in our corpus.

Highlights

In recent years, evaluations of normalization methods for vowel formant data have prioritized the maximization of vowel-space similarity (e.g., Adank, Smits, & Van Hout, 2004; Flynn & Foulkes, 2011)
The a priori preference for normalization methods that maximize the similarity of normalized vowel spaces begs the question of the similarity of the vowel spaces and may erase important phonetic, within-category variation from our data
The limiting case would be a “saturated model” that completely erases within-category variation, collapsing all productions of each phoneme into a single point in the normalized space. The output of such a method would not be useful for most researchers, suggesting that some constraints on the power of normalization methods are necessary in practice

Summary

Introduction

Evaluations of normalization methods for vowel formant data have prioritized the maximization of vowel-space similarity (e.g., Adank, Smits, & Van Hout, 2004; Flynn & Foulkes, 2011). The a priori preference for normalization methods that maximize the similarity of normalized vowel spaces begs the question of the similarity of the vowel spaces and may erase important phonetic, within-category variation from our data. This suggests that there should be perceptual constraints on normalization methods: the ideal method will not remove all within-category variation but just the variation that is perceptually removed by listeners From this perspective, it is possible to “overnormalize” (Barreda & Nearey, 2018) vowel spaces by removing variation that listeners do not remove in perception, resulting in the removal of legitimate phonetic variation from a dataset

Methods

Results

Discussion

Conclusion