Evaluation of vowel normalization procedures.

Abstract

Vowel normalization procedures are commonly evaluated on the basis of how effectively they separate the vowels of a single test data set into distinct groups corresponding to the phonetic categories of that language. A quantifiable method of evaluation is proposed here, based on how much of the overall variance is removed from the data. This evaluation method is applied to the vowels of six different Germanic languages which have been normalized according to four different procedures. It is shown that no one normalization procedure is the most effective for all languages. Furthermore, some of the most successful of these normalizations introduce procedural artifacts into the data, and as a result the relative quality of vowels across languages or dialects is altered. In such cases, it is shown that comparison of the normalized vowels of one language with the (separately) normalized vowels of another language are not valid if the vowel systems are different. Some reasons for the appearance of these procedural artifacts are discussed.