PhD defence, Pierre-Alexandre Broux

Date : 10/01/2020
Time : 14h00
Location : Room 210, IC2 building, LIUM, Le Mans Université

Title : Speaker diarization in audiovisual files in interaction with human annotators

Jury members :
– Jean-François BONASTRE (LIA, Université d’Avignon)
– Nicholas EVANS (EURECOM)
– Régine ANDRE-OBRECHT (Université Toulouse 3)

– Sylvain MEIGNER (LIUM, Le Mans Université)

Co-supervisors: :
– Simon PETITRENAUD (LIUM, Le Mans Université)


Abstract :

The diarization task tries to determine the number of speakers as well as their interventions in an audio file. It is an interesting task for any enterprise willing to index its audiovisual contents. Especially, the French National Audiovisual Institute (INA) desires to apply this task on its archives so as to improve its accessibility and its annotation.
However, the uses of the institute need a minimal quality which, most of the time, is not reached by the state-of-the-art automatic diarization systems yet. In order to reach the wanted effectiveness, a human can correct the output of a diarization system. Nevertheless, a human intervention is generally time-consuming and expensive.

In order to reduce these costs, a possible solution is to use a computer-assisted a human gives some information to a system in order that it can improve its predictions so as to decrease its intervention cost. The present manuscript revolves around the computer-assisted diarization. It proposes a metric so as to assess the human intervention cost to correct a diarization, a framework to evaluate the human corrections of a speaker diarization, an automaton simulating the human corrections to do for a diarization and some computer-assisted diarization systems decreasing the total human intervention cost. More precisely, the proposed computer-assisted diarization systems reassess either only the speaker clustering or the segmentation and the speaker clustering.

Keywords :
Diarization, Computer-assisted system, Human-computer interaction (HCI), Annotation