Project WISE is working to develop a measurement tool that can measure teachers’ implementation of evidence-based behavior management practices in general and special education classrooms. Typically, researchers and school administrators will use systematic direct observation (SDO) to collect data on teacher practices; however, this can be resource intensive and requires rigorous observer training. The rapid development of wearable technologies and machine learning offer an alternative, low-cost method for collecting these data in schools. We envision using low-cost wearable sensors included in modern smart watches to collect classroom audio data and using deep learning models to code and quantify teacher practices.

Last semester, our team focused on a few initial teacher behaviors (i.e., praise statements, reprimands, opportunities to respond) and began training two wav2vec models. One model is focusing on using fixed-length audio segmentation and a multi-label model. The other model is focusing on using pyannote to partition the classroom audio into segments and using a single label model. This semester, our team will work to finish the initial training of these two models. Then we will analyze and compare the performance of the two models to determine feasibility, accuracy, and preference.