CHATTR: An open-source tool for estimating turn-taking behavior

• In Event: Poster Session 04
  In Poster Session: PS 04 Topic - Language, Communication

Abstract

While children around the world can and do acquire their home languages with little directed speech in the first few years, more frequent child-caregiver turn-taking can indicate a pedagogically rich language environment that may contribute positively to vocabulary growth and other behaviors normatively associated with literate language development (e.g., Hirsh-Pasek et al., 2015; Romeo et al., 2018a, 2018b; Rowe 2012; Warlaumont et al., 2014; Pretzer et al., 2019; Yoo et al., 2018). Spontaneous home turn-taking behaviors may therefore be useful for predicting individual differences in later language outcomes.

The conventional use of turn-taking measures in language development research has been boosted by the LENA system—a hardware and software package that makes simple work of audio-recording and automatically analyzing children’s at-home environments. Among the system’s automated measures is a Conversational Turn Count (CTC), which estimates the number of turn-taking exchanges between the target child and a nearby adult interactant based on the proximity of their utterances in time. While CTC is prone to error (Bergelson et al., 2020), it correlates with language outcome measures (e.g., Romeo et al., 2018) and makes possible the large-scale analysis of child language environments.

The LENA system, however, has a number of drawbacks: First, its timing constraints are much longer than what is found in manual studies of parent-child interaction, and it is therefore unclear what is being detected (Hilbrink et al., 2015; Casillas et al., 2016). Second, the system’s simplified interface makes it difficult for researchers to execute alternative turn-taking analyses without significant scripting expertise. Third, researchers using non-LENA recording devices are unable to use the software. Finally, it is expensive—an obstacle in promoting scientific inquiry across more culturally, linguistically, and economically diverse contexts.

We introduce an open-source alternative software package (chattr, an R package) for detecting and characterizing turn-taking behavior in child-centric recordings of any type. Chattr works on utterance annotations, which can be generated by a human annotator (e.g., in ELAN), the LENA system, or a third party diarization tool (e.g., an open-source tool). Chattr, like LENA, uses temporal contingency to detect turn-taking behaviors, but yields a rich report on the features of individual turn transitions and turn-taking bouts, including their timing, and can be tailored to suit turn-taking-related analyses of many types. The system also estimates the expected baseline rate of contingent talk; useful for establishing that the turn-taking behaviors found exceed what is expected by chance.

We demonstrate use of this package by sharing preliminary results of a study describing the turn-taking environments of 70 children under age 3;0 growing up across four linguistic populations for which we have human-annotated utterance information (North America, Argentina, Mexico, and Papua New Guinea). We compare the turn-taking measures to parallel findings on these populations with child-directed speech frequency, addressing similarities and differences in the measures.

Author

• Marisa Casillas, University of Chicago
  Presenting Author