Measuring Prosodic Entrainment in Conversation: A Review and Comparison of Different Methods

This study aims to further our understanding of prosodic entrainment and its different subtypes by analyzing a single corpus of conversations with 12 different methods and comparing the subsequent results. Entrainment on three fundamental frequency features was analyzed in a subset of recordings from the LUCID corpus (Baker & Hazan, 2011) using the following methods: global proximity, global convergence, local proximity, local convergence, local synchrony (Levitan & Hirschberg, 2011), prediction using linear mixed-effects models (Schweitzer & Lewandowski, 2013), geometric approach (Lehnert-LeHouillier, Terrazas, & Sandoval, 2020), time-aligned moving average (Kousidis et al., 2008), HYBRID method (De Looze et al., 2014), cross-recurrence quantification analysis (e.g., Fusaroli & Tylén, 2016), and windowed, lagged cross-correlation (Boker et al., 2002). We employed entrainment measures on a local timescale (i.e., on adjacent utterances), a global timescale (i.e., over larger time frames), and a time series-based timescale that is larger than adjacent utterances but smaller than entire conversations. We observed variance in results of different methods. Results suggest that each method may measure a slightly different type of entrainment. The complex implications this has for existing and future research are discussed.