Measuring the speech level and the student activity in lecture halls: Visual- vs blind-segmentation methods

•The blind-segmentation methods were used to measure speech level and student activity during thirteen lessons.•Gaussian Mixture Model seems to be the most robust technique K-means clustering seems to overestimate the speech level, due to short-time integration.•Percentile-level based techniques could be improved basing on analytical considerations Inverse Lombard effect was found in the students? behaviour. The background noise has a fundamental role in oral communication, since the higher the speech level with respect to the background noise the greater the intelligibility. In occupied lecture halls the main contribution to background noise is related to the human noise, which is called by scholars student activity. Scholars proposed methods to measure both student activity and speech level through short-time sound level meter measurements during lessons. However, a comparison of their relative effectiveness on a relevant set of data in different situations is still lacking. In this study, basing on recordings of university lessons performed with public address system, student activity and speech level values were extracted using different methods. Various scenarios of university lectures were recorded: frontal lessons, media-aided lectures, open discussions. Visual-segmentation and blind-segmentation procedures were compared for each case. Results show the benefits of blind-segmentation methods, which seem to be reliable and affordable methods for this kind of analyses.