Using acoustic string-edit distance to evaluate US pronunciation variation
Clelia R. LaMonica · The Journal of the Acoustical Society of America · 2014
String-edit distances, such as the Levenstien distance, have been used in perceptual-linguistic studies to compare differences among language varieties, based on changes that occur between two speech samples (Gooskens and Heeringa, 2004; Heeringa et al., 2006; Nerbonne et al., 2008). However, this method relies mainly on phonetic transcriptions rather than actual speech data. This work illustrates how acoustic measurements are taken from speech samples from across the United States, and a distance measurement between them is derived for use in further perceptual comparisons (such as perceived distance from standard, intelligibility, appeal, and identification). The speech samples consist of six sentences from various regions of the United States, each sentence containing phonological features that may be marked as perceptually significant for dialect identification (Clopper 2011, Labov et al. 2005, Thomas 2001). The methodology for assessing phonetic distance between two regional varieties is addressed, in particular, by using the Euclidean distance between normalized Bark formants of phonological features within the samples. The benefits and potential disadvantages to using acoustic data vs. transcribed data are addressed as well.