Active selection with label propagation for minimizing human effort in speaker annotation of TV shows

In this paper an approach minimizing the human involvement in the manual annotation of speakers is presented. At each iter- ation a selection strategy choses the most suitable speech track for manual annotation, which is then associated with all the tracks in the cluster that contains it. The study makes use of a system that propagates the speaker track labels. This is done using a agglomerative clustering with constraints. Several dif- ferent unsupervised active learning selection strategies are eval- uated. Additionally, the presented approach can be used to ef- ficiently generate sets of speech tracks for training biometric models. In this case both the length of the speech track for a given person and its purity are taken into consideration. To evaluate the system the REPERE video corpus was used. Along with the speech tracks extracted from the videos, the op- tical character recognition system was adapted to extract names of potential speakers. This was then used as the 'cold start' for the selection method.