Skip to Main content Skip to Navigation

Analyse et détection des émotions verbales dans les interactions orales

Abstract : The thesis addresses the representation and automatic detection of emotions in natural speech. Most experiments were conducted on data recorded in a 20-hour real-life human-human call center. In a first part, we present and validate an annotation scheme allowing the annotation of emotion mixtures. Several annotations are combined in an "emotion vector" and the presence of many blended emotions is thus revealed. Those emotion mixtures are further studied with two perceptive tests. In the second part, more than a hundred paralinguistic cues are extracted per emotion segments and the non complex segments are used to train classifiers, mostly Support Vector Machine. Discrimination experiments are done with 2 to 5 emotion classes. Some take into consideration the speaker gender and role, i.e. agent vs. client. The relative importance of different paralinguistic cues as well as the combination of linguistic and paralinguistic cues are also studied. In addition, during a collaboration between different sites involved in the HUMAINE network of excellence, we have been able to compare and combine our expertise on a common corpus. The results obtained by LIMSI were at the state of the art. Finally, we study the performances of classifiers trained and tested on different corpus. In the case of acted speech and natural speech, models trained on one type of data do not necessarily work on the other type of data.
Document type :
Complete list of metadatas
Contributor : Magali Roserat-Brilhac <>
Submitted on : Thursday, September 15, 2011 - 4:51:47 PM
Last modification on : Thursday, December 10, 2020 - 12:30:35 PM
Long-term archiving on: : Friday, December 16, 2011 - 2:27:13 AM


  • HAL Id : tel-00624085, version 1



Laurence Vidrascu. Analyse et détection des émotions verbales dans les interactions orales. Informatique [cs]. Université Paris Sud - Paris XI, 2007. Français. ⟨tel-00624085⟩



Record views


Files downloads