Impact of weak labels for ambient sound analysis

In the domain of ambient sound analysis, many applications are dealing with unlabeled or weakly labeled data as opposed to strongly labeled data. Weak labels indicate which sound events happened in an audio clip. Strong labels indicate which sound events happened in an audio clip and when (with timestamps). Multiple approaches in the literature have been proposed to solve the problem of weak labels since most datasets in the community are labeled with weak labels, but in a real scenario also involving multilabels, overlapping sounds, unbalanced classes, it is difficult to interpret the results and state that the problem solved is indeed the weakly labeled data problem.

We propose a way to isolate the problem of weakly labeled data by creating a dataset, and then, we analyse the impact of different solutions proposed (from the litterature and new ones) in our audio tagging system within different scenarios involving different granularity of weak labels.