Signal-to-noise ratio in university lecture halls with low intelligibility

Speech intelligibility plays a key role in determining the quality of verbal communication. It depends on the acoustic characteristics of the room and the signal-to-noise ratio (SNR). Background noise has three main components: HVAC noise, anthropic noise (student activity) and external activities noise, each of them with different noise spectra. In this work, measurements taken with a sound level meter during lectures in university classrooms are analyzed using advanced statistical techniques in order to detect and separate the various components which contribute to background noise. The same statistical techniques are used to characterize the speech level too, i.e. the signal in the SNR. From the raw data collected with a sound level meter, an asymmetrical distribution is built. Then four statistical techniques are applied: percentile levels, Gaussian mixture model based on peak detection, blind Gaussian mixture model and blind k-means clustering. Results are compared and discussed, highlighting the pros and cons of each technique.