Early frame-based detection of acoustic scenes

Let us consider a specific acoustic scene appearing in a continuous audio stream recorded while making a trip a in city. In this work, we aim at detecting at the earliest opportunity the several occurrences of this scene. The objective in early detection is then to build a decision function that is able to go off as soon as possible from the onset of a scene occurrence. This implies making a decision with an incomplete information. This paper proposes a novel framework in this area that i) can guarantee the decision made with a partial observation to be the same as the one with the full observation; ii) incorporates in a non-confusing manner the lack of knowledge about the minimal amount of information needed to make a decision. The proposed detector is based on mapping the temporal sequences to a landmarking space thanks to appropriately designed similarity functions. As a by-product, the built framework benefits from a scalable learning problem. A preliminary experimental study provides compelling results on a soundscape dataset.