A comparison of algorithms to automatically extract vowel formants and nasal poles from tagged vowels in PRAAT.

Abstract

The goal of the present research was to develop an easy-to-use PRAAT script yielding robust measurements of acoustic correlates of vowel quality and vowel nasality. Several algorithms were scripted in PRAAT and evaluated on two corpora. The first was the Hillenbrand corpus, containing /hVd/ words in isolation, spoken by 50 men, 50 women, 29 boys, and 21 girls, where vowels have been hand-tagged by experts. The best performing algorithm yielded extremely high correlations with the expert-measured F1 and F2 frequencies (r>0.95). The second corpus consisted of spontaneous speech by American English and Canadian French women, addressing either their infant or an adult, and where several types of vowels had been hand-tagged (including point vowels /i,a,u/, frequently used to assess vowel space size; and contrasts of vowel tenseness, likely involving F1 and F2 frequencies; and vowel nasality, correlated with F1 bandwidth, as well as its amplitude in relation to the nasal poles P0 and P1). Although, even the best algorithm yielded spurious F2 values for infant-directed /u/'s, measures were robust indicators of vowel quality and nasality in non-back vowels across the different speakers, languages, and registers. Thus, this script may be useful, particularly for non-specialists, since it only requires vowel-tagging.