During an interaction, interactants exchange speaking turns. Exchanges can be done smoothly or through interruptions. Listeners can display backchannels, send signals to grab the speaking turn, wait for the speaker to yield the turn, or even interrupt and grab the speaking turn. Interruptions are very frequent in natural interactions. To create believable and engaging interaction between human interactants and embodied conversational agent ECA, it is important to endow virtual agent with the capability to manage interruptions, that is to have the ability to interrupt, but also to react to an interruption. As a first step, we focus on the later one where the agent is able to perceive and interpret the user's multimodal behaviors as either an attempt or not to take the turn. To this aim, we annotate, analyse and characterize interruptions in humanhuman conversations. In this paper, we describe our annotation schema that embeds different types of interruptions. We then provide an analysis of multimodal features, focusing of prosodic features (F0 and loudness) and body (head and hand) activity, to characterize interruptions.