Acoustic correlates of accentedness and comprehensibility in L2 Spanish judgments: A mixed-effects modeling approach

Abstract

Two constructs—comprehensibility and accentedness—figure themselves prominently in listeners’ judgments of L2 (second language) speech. Correlational analyses have shown that they make separate contributions to such judgments (Derwing and Munro, 1995, 1997, 2005). As it is important to set realistic goals for adult L2 learners by prioritizing understanding over nativelikeness (Levis, 2005), recent studies began to examine the relative contribution of linguistic aspects, especially acoustic features, to comprehensibility and accentedness (Munro & Derwing, 1999; Kang, Rubin, & Pickering, 2010; Trofimovich & Issac, 2012). Consistent with this agenda, this study first examines the reliability of scores (produced by the 9-point Lickert scale) under modern measurement framework (e.g. multi-facet Rasch measurement model). Then, the relationship between acoustic features, including the confidence values (for each utterance) obtained using Google Cloud Platform speech recognition engine as well as suprasegmental features extracted using Prosogram (Mertens, 2014), and ratings are established using mixed-effects modeling techniques. Results indicate that, even though it can be shown that comprehensibility and accentedness are statistically distinct constructs, both ratings share a common set of acoustic correlates as significant predictors—speech time, standard deviation of pitch values, and normalized pairwise variability index (nPVI).