Shout analysis and characterisation

The goal of this paper is to analyse and to characterise the shout of a people to develop an automatic detector. We define a shout as a voiced part of an audio signal maintained over time. We show that a set of formants parameters can be defined to discriminate a typical “shout” from a “neutral” part of a speech. Moreover, it appears clearly that the duration of the window used to estimate these parameters is critical to yield better results. We conclude by presenting a performance analysis in the noisy context of a transport surveillance application.