The role of principal component analysis in vowel acquisition research

Abstract

Traditional formant analysis techniques have yielded questionable results, due to a high fundamental frequency in young children’s speech. Therefore, an alternative approach for researching vowel acquisition appears desirable. A band-filtering analysis method was developed that minimizes the dependence of the results on F0 to measure the spectral envelopes in children’s utterances automatically. In PRAAT scripts, first the criteria are set with regard to sound quality (to avoid clipped fragments) and F0 (to avoid vowel fragments with F0 above 425 Hz). Secondly, the spectral envelopes covering a range from 0 to 7 kHz, resulting in 40 filter values per spectrum, are calculated. A matrix is produced, representing intensity in each of the 40 filters. Data reduction is achieved via principal component analysis. Previous research (Pols, 1977) suggests that the first two principal components are related to the F1 and F2 values. Analyses of unlabeled vowel productions in 2-year-old Hungarian- and Dutch-speaking children show that the first two eigenvectors account for 55% of the variability. By using these two eigenvectors, a reference plane is created. Mapping labeled vowel measures onto the plane suggests that children produce language-specific vowel categories at this young age. [Work supported by UW and UvA/ACLC.]