An Automatic Teacher Activity Recognition System based on Speech Transcriptions

The importance of classroom observation as a tool to improve the quality of education has motivated the use of Machine Learning methods for automatic detection of classroom activities. However, most of these efforts require expensive sensors to register video and/or audio signals from lessons or do not follow a validated observation protocol. In contrast, in this work we present a cost-effective, easy to use, non-intrusive scalable method designed for the needs and practical conditions of teachers. We recorded teacher’s talk relying only on a smartphone with a budget lavalier microphone and followed a validated observation protocol: the Classroom Observation Protocol for Undergraduate STEM (COPUS). Using teachers’ talk transcriptions of 41 fourth grade lessons, we trained Machine Learning models to recognize three collapsed categories from COPUS’ teacher dimension throughout a lesson (Presenting, Guiding, and Administration). In particular, we propose a deep network model which outperformed a powerful model like BERT for this task. With this method we contribute with a useful tool to provide fast and effective feedback for teachers, as well as the possibility for researchers to quickly analyze in a thoughtful way a large amount of lessons.