Analysis of mechanical and human activity noise in offices using machine learning algorithms

International audienceTask performances of workers are tightly tied to the com-fort of the environment. This depends on multiple factorssuch as spatial, lighting, air quality, thermal and acousticconditions. Noise is one of the most impacting contributorsof annoyance. Offices are affected by two main sourcesof background noise: the mechanical sources and the hu-man activity, in addition the traffic noise, if present. Inmany cases an high mechanical noise, attributable to of-fice devices (computers, phones, etc...), HVAC systems,can trigger the Lombard effect among people and, con-sequently, the noise due to the human activity increasesas well as all related issues. Speech is detected as one ofthe most annoying sound sources [1]. ISO standard, con-cerning the acoustics of open plan offices, indicates thatSpeech Transmission Index (STI) has to be measured inunoccupied condition. This results in neglecting the hu-man activity component on tasks performance. STI basi-cally depends on two factors: the reverberation time andthe background noise. Since the human noise represents adynamic source, its evaluation becomes essential. Cluster-ing algorithms, largely used in machine learning as well,were employed in classrooms in order to detect the humannoise due to the students during lessons [2, 3] and open-plan offices [4