Date: 02/06/2023
Heure: 11h00
Lieu: IC2, salle des conseils
 
 
DEFT 2023: Mohamed Ettaleb, Clément Besnard, Nathalie Camelin

Our paper describes the participation of the LIUM-IRISA team in the DEFT 2023 evaluation campaign. Our team participated in the main task, which this year consists of developing approaches for automatically answering medical multiple-choice questions. We have implemented several systems, the first use a knowledge base, a second use generative model-based system, and a final system combining a set of descriptors.

 
Blizzard 2023: Félix Saget, Thibault Gaudier, Meysam Shamsi, Marie Tahon

Our team participated in the annual speech synthesis challenge, which intend to foster the development of Text to Speech (TTS) systems.This year, the challenge proposed two tasks. The first task aims at creating a French voice with a large speech dataset (around 50 hours) from one single speaker. The second task aims at adapting an existing TTS to a new voice with few speech data (around 2 hours). Our choice lies on a mono speaker TTS based on FastSpeech trained on the large dataset, combined with a voice conversion based on Tacotron2 trained on the small dataset. The perceptual evaluation step is under progress, and we will present you the different aspects that are evaluated.