Multi-microphone acoustic events detection and classification for indoor monitoring

Digital processing of the information provided by the real world is fundamental for designing automated solutions for daily tasks. The complex information of real acoustic scenes must be processed to be used for further applications. This information is composed of noise and audio events. In the current paper, a system to classify seven indoor acoustic events is analysed. The detection and classification algorithm is based on the baseline algorithm of DCASE'2016 challenge. The acoustic features consist of MFCC coefficients, delta coefficients and acceleration coefficients. The system includes binary GMM based classifier for each sound event class. Different microphone positions have been tested. Two of these configurations are composed of four microphones, and the other with one microphone. The results of simulations predict a better classification when more than one microphone is used, that is, more than one classifier. This approach, combined with majority voting techniques, reduces classification errors for any acoustic condition. An increase in system robustness against reflection coefficient changes is observed too. Moreover, the results with multi-microphone configurations confirm the improvement, covering a bigger part of space at the same time.