
LEAD Lab language environments
Built the survey-delivery system for the SLIDE study and a Python package that computes speech burstiness from day-long LENA home recordings.
The problem
SLIDE, a study at the Language, Experience, and Development Lab, asks how real-time caregiver stress and a preschooler's language environment shape development. Answering it takes two streams of data: stress reports collected from caregivers several times a day, and day-long home audio from LENA recorders. The scientific purpose is to see how stress and language input move together within a day. My contribution is the software that collects the first stream and measures structure in the second.
What I did
I built the ecological momentary assessment system: a Python generator for randomized survey times, and automated delivery through the REDCap API so each study event sends its own survey link. For the audio, I wrote a Python package that turns LENA files into 5-minute epochs, keeps awake child-directed speech, and computes burstiness, a measure of how speech clusters into dense periods across a day. I am also building iPad tools for administering child assessments, and I am trained to administer the CELF, Bracken-3, and Matrix Reasoning.
How it works
The burstiness pipeline follows the method of Cychosz, Romeo, Edwards, and Newman (2024): parse LENA 5-minute sections, join sleep and child-directed-speech labels, drop recordings of 5 hours or less, interpolate the adult word count series, and mark a burst wherever it exceeds three standard deviations. It runs from a config file or the command line and has unit tests, and its epoch parser matches the published epoch tables across 194 recordings to floating-point precision. Raw data stays out of version control because the recordings identify children.
What happened
SLIDE has a working survey system ahead of its first participants, and the lab has a burstiness pipeline it can rerun on any LENA corpus with one command. I also set up the lab GitHub organization and proposed immutable raw-data ingests, so the next analysis starts from shared, versioned code.