The impact of automated phonetic alignment and formant tracking workflows on sound change measurement
C. I. Watson, Z. E. Evans, Elaine Ballard, B. Ross, Steve Cassidy · 2024
In this study we investigated the impact that forced aligners and formant trackers have on sound change measurements. We compared the results from handchecked data to those where the process was automated. Data was taken from read passages in the Auckland Voices corpus of New Zealand English (NZE). There were 60 speakers: 33 under 25 years (17 women, 16 men) and 27 over 40 years of age (16 women and 11 men). Three forced aligners (MAUS, FAVE aligner, and Montreal Forced Aligner (MFA)) and three formant trackers (FOREST, SNACK, and PRAAT) were investigated in five different combinations (MAUS with FOREST, SNACK, and PRAAT and additionally FAVE and MFA with PRAAT). The formant values obtained using these processes were compared with formant data from hand-checked vowel tokens (both vowel boundaries and formant tracks). The hand-labelled data only used vowels with phrase stress. The study looked at the movements of the short front vowels in NZE and measured these changes using the vowel space measure. The results showed that there was significant variability introduced into an analysis of sound-change based on an automated forced-alignment and formant extraction workflow. As such, to support reproducibility of studies, it is critically important to include enough details about this workflow to avoid any ambiguity about the process.