Autoapprentissage pour le regroupement en locuteurs : premières investigations (First investigations on self trained speaker diarization )
1767Authors
nan
Abstract
AbstractThis paper investigates self trained cross-show speaker diarization applied to collections of French TV archives, based on an i-vector/PLDA framework. The parameters used for i-vectors extraction and PLDA scoring are trained in a unsupervised way, using the data of the collection itself. Performances are compared, using combinations of target data and external data for training. The experimental results on two distinct target corpora show that using data from the corpora themselves to perform unsupervised iterative training and domain adaptation of PLDA parameters can improve an existing system, trained on external annotated data. Such results indicate that performing speaker indexation on small collections of unlabeled audio archives should only rely on the availability of a sufficient external corpus, which can be specifically adapted to every target collection. We show that a minimum collection size is required to exclude the use of such an external bootstrap.
Keywords
Full Content
Autoapprentissage pour le regroupement en locuteurs : premi`eres investigations Ga¨el Le Lan1,2 Sylvain Meignier1 Delphine Charlet2 Anthony Larcher1 (1) LIUM, Universit´e du Maine, France (2) Orange Labs, France prenom.nom@lium.univ-lemans.fr, prenom.nom@orange.com R´ESUM´E Cet article traite de l’autoapprentissage d’un syst`eme i-vector/PLDA pour le regroupement en locuteurs de collections d’archives audiovisuelles franc¸aises. Les param`etres d’extraction des i-vectors et du calcul des scores PLDA sont appris de fac¸on non supervis´ee sur les donn´ees de la collection elle-mˆeme. Diff´erents m´elanges de donn´ees cibles et de donn´ees externes sont compar´es pour la phase d’apprentissage. Les r´esultats exp´erimentaux sur deux corpora cibles distincts montrent que l’utilisation des donn´ees des corpora en question pour l’apprentissage it´eratif non supervis´e et l’adaptation des param`etres de la PLDA peut am´eliorer un syst`eme existant, appris sur des donn´ees annot´ees externes. De tels r´esultats indiquent que la structuration automatique en locuteurs de petites collections non annot´ees ne devrait reposer que sur l’existence d’un corpus externe annot´e, qui peut ˆetre sp´ecifiquement adapt´e `a chaque collection cible. Nous montrons ´egalement qu’une collection suffisamment grande peut se passer de l’utilisation de ce corpus externe. ABSTRACT First investigations on self trained speaker diarization This paper investigates self trained cross-show speaker diarization applied to collections of French TV archives, based on an i-vector/PLDA framework. The parameters used for i-vectors extraction and PLDA scoring are trained in a unsupervised way, using the data of the collection itself. Performances are compared, using combinations of target data and external data for training. The experimental results on two distinct target corpora show that using data from the corpora themselves to perform unsupervised iterative training and domain adaptation of PLDA parameters can improve an existing system, trained on external annotated data. Such results indicate that performing speaker indexation on small collections of unlabeled audio archives should only rely on the availability of a sufficient external corpus, which can be specifically adapted to every target collection. We show that a minimum collection size is required to exclude the use of such an external bootstrap. 1 Introduction La tˆache de Segmentation et Regroupement en Locuteurs (SRL) vise `a ´etiqueter les locuteurs dans un ou plusieurs enregistrements audios, sans connaissance a priori des locuteurs. L’augmentation constante des volumes de donn´ees n´ecessite une indexation efficace. La SRL appliqu´ee `a des collec- tions est une tˆache globale qui consiste `a traiter un ensemble d’enregistrements audios bruts (non segment´es en tours de parole) afin d’identifier de mani`ere unique les tours de paroles de chaque locuteur. Cette tˆache se d´ecompose g´en´eralement en deux ´etapes : la SRL intra-enregistrement, o`u il Actes de la conférence conjointe JEP-TALN-RECITAL 2016, volume 1 : JEP 82 s’agit de segmenter et regrouper les occurrences des locuteurs au sein d’un mˆeme enregistrement, et le regroupement inter-enregistrements, qui vise `a regrouper les locuteurs intra-enregistrements (Meignier et al., 2002). D’autres impl´ementations sont possibles, o`u tous les enregistrements peuvent ˆetre concat´en´es en un super-enregistrement, pour ensuite ˆetre trait´es comme un probl`eme artificiel de SRL intra-enregistrement (Tran et al., 2011). L’approche en deux ´etapes, plus naturelle, est la plus couramment …
Content truncated to 500 words
Labeled Text
AbstractThis paper investigates self trained cross-show speaker diarization applied to collections of French TV archives, based on an i-vector/PLDA framework. The parameters used for i-vectors extraction and PLDA scoring are trained in a unsupervised way, using the data of the collection itself. Performances are compared, using combinations of target data and external data for training. The experimental results on two distinct target corpora show that using data from the corpora themselves to perform unsupervised iterative training and domain adaptation of PLDA parameters can improve an existing system, trained on external annotated data. Such results indicate that performing speaker indexation on small collections of unlabeled audio archives should only rely on the availability of a sufficient external corpus, which can be specifically adapted to every target collection. We show that a minimum collection size is required to exclude the use of such an external bootstrap.
Publication Details
- Paper ID:
- 1767
- Year:
- 2016
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan