Nelly Barbot
IRISA / Université de Rennes 1
Session JEP poster P2 Lundi 9 Juin - 16h00 18h00
-
papier 1642
Evaluation de méthodes de réduction de corpus linguistiques
- Nelly Barbot ( IRISA / Université de Rennes 1 - ENSSAT, Lannion)
- Pierre Alain ( IRISA / Université de Rennes 1 - ENSSAT, Lannion)
- Olivier Boeffard ( IRISA / Université de Rennes 1 - ENSSAT, Lannion)
- Jonathan Chevelu ( IRISA / Université de Rennes 1 - ENSSAT, Lannion)
- Arnaud Delhay ( IRISA / Université de Rennes 1 - ENSSAT)
- Résumé : This article deals with covering methodologies in the context of automatic speech processing technologies. More precisely, we are interested in covering phonological attributes of a linguistic corpus under the constraint of a minimal duration. This goal is classically achieved by greedy algorithms which however do not guarantee the optimality of the solutions. We propose to compare the results of a new algorithm, the LamSCP, that calls upon the principles of lagrangian relaxation, and an agglomeration-spitting greedy algorithm to achieve an optimal covering. We conducted experiments on the Gutenberg corpus considering, phone, diphone and triphone optimal covering. The LamSCP provides better solutions than the greedy algorithm and enables to locate their quality by offering a lower bound to the optimization problem.
- article
Session JEP poster P6 Jeudi 12 Juin - 10h30 12h30
-
papier 1598
Transformation de la prosodie par adaptation MLLR de GMM
- Damien Lolive ( IRISA / Université de Rennes 1)
- Nelly Barbot ( IRISA / Université de Rennes 1)
- Olivier Boëffard ( IRISA / Université de Rennes 1)
- Résumé : In a voice transformation context, prosody transformation using parallel corpora is quite unrealistic as such corpora are difficult and also expensive to build. Based on this observation, we propose an approach for transforming prosody using non-parallel corpora thanks to the MLLR adaptation strategy. This methodology is applied to the joint transformation of duration and F0 at the syllable level. The source data are modelled by a GMM which is adapted to the target by applying a linear transformation to the mean vectors of the gaussian mixture. This methodology is applied to the conversion of duration and F0 between two french speakers and is evaluated by cross validation between the models and the test datasets.
- article