Perceptual validation of vowel normalization methods for variationist research
Researchers often use vowel-normalization methods to minimize between-speaker variation in production data. Typically, these methods are evaluated based on their ability to maximize the similarity the vowel spaces of different speakers. Unfortunately, this approach may favor methods that ‘overnormalize’ data, removing perceptible phonetic variation that may be linguistically meaningful. Instead, since linguists use normalized data to infer differences in vowel quality, which exists only in the minds of human listeners, we should consider the likelihood that different normalization methods will reflect listener judgments. From this perspective, methods that maximize the similarity of normalized data but do not reflect the judgments of human listeners are not useful for most purposes in linguistic research. Based on the poor theoretical support for Lobanov normalization and its poor statistical properties, we suggest that the Lobanov method is too powerful, demonstrably removes phonetic variation from data, and should be avoided for variationist research.
