Poster #205 - Emotion in the Language Environment: Assessing LENA’s Detection of Toddlers’ Vocal Negative Emotion

• In Event: 1-005 - Poster Session 01
  In Poster Session: PS 01 Section - Social, Emotional, Personality

Integrative Statement

Children communicate with both emotions and language, and parents’ responses to emotional and verbal cues are critical for development. Yet emotional and verbal communication are often studied separately (Cole, Armstrong, & Pemberton, 2010). In the toddler period, when negative emotion peaks and verbal expression is emerging (Brownell & Kopp, 2010), it is particularly important to consider both emotional and verbal expression. We present a pilot study that examined the utility of Language Environment Analysis (LENA; Xu, Yapanel, & Gray, 2009), an automated processing tool, for studying emotional and verbal communication in toddlers.

LENA audio-records and quantifies young children’s language environment, generating estimates of language input and output. An underutilized feature of LENA is its identification of vocal negative emotion expressions (e.g., fusses, whines, cries), which we refer to as cries. LENA’s accuracy in identifying adult speech and child verbal/preverbal vocalizations is established (Xu et al., 2009) but its cry detection accuracy has not been studied.

A LENA recorder sampled the audio environment of twenty-five 1-year-olds (12-23 months, Mean=16.60 months; 60% female) for a full day. LENA’s software provided continuous assessment of adult speech, toddler verbal/preverbal vocalizations, toddler cries, and other sounds. As advised by the LENA Foundation, we randomly selected 30% of participants (N=8), and isolated three 10-minute sections identified by LENA as being in the top and bottom 10th, and middle 20th, percentile for cry frequency for each participant. A trained research assistant listened to these sections and identified onsets and offsets of toddler cries.

We calculated agreement estimates between human-identified and LENA-identified cries by generating a confusion matrix based on match or mismatch of 10ms frames (Table 1). LENA detected 79% of frames identified by the human coder as cries, comparable to previous estimates of LENA’s vocalization detection (75%) and adult speech detection (82%). LENA’s precision, however, was lower; of the frames identified by LENA as a cry, the human coder agreed 58% of the time. The overall agreement between human and LENA cry detection was acceptable (Kappa=.67); moreover, total cry duration estimates for human-coded and LENA files and files were highly correlated, r(8)=.787, p=.021.

LENA’s algorithm classifies audio signals into mutually exclusive categories, such that it does not generate codes for a child speaking and crying. The RA further identified instances of combined vocalization/cry (see Table 1). Only 9% of LENA-identified cries, and 5% of LENA-identified vocalizations, were classified by the human coder as a vocalization/cry, consistent with research suggesting that toddlers have difficulty speaking while expressing negative emotion (Bloom & Beckwith, 1989). Thus, the majority of LENA-identified cries were mutually exclusive, and did not contain verbal content.

In sum, LENA demonstrates adequate agreement with human coders in its identification of toddlers’ vocal negative emotion expressions, although it may over-identify cries. We will also report on potential sources of variability in human/LENA agreement, including child sex, age, language ability and temperament. We discuss he benefits and drawbacks of ambulatory automated assessment, and implications for the study of the intersection of early emotional and verbal communication and development.

Authors

• Margaret A Fields-Olivieri, The Pennsylvania State University
  Presenting Author

• Sarah S. Shahriar, The Pennsylvania State University
  Non-Presenting Author

• Pamela M. Cole, The Pennsylvania State University
  Non-Presenting Author