Début du message réexpédié :De: Emmanuel Dupoux <emmanuel.dupoux@gmail.com>Objet: [ZR21] Announcing the Zero Resource Challenge 2021: spoken language modeling from raw audioDate: 25 novembre 2020 à 18:51:52 UTC+1Cc: Ewan Dunbar <ewan.dunbar@gmail.com>, Mathieu Bernard <mathieu.a.bernard@inria.fr>, Maureen de S <maureen.deseyssel@gmail.com>, Tú Anh Nguyễn <nguyentuanh208@gmail.com>, Nick Hamilakis <nick.hamilakis562@gmail.com>Dear Colleague,We have the pleasure to announce the new iteration of the Zero Resource Speech Challenge, which is now open! It will be submitted as a challenge session to Interspeech 2021. Its aim is to build a language model from raw audio without any text or phonetic labels (Language Modelling from Speech). We take inspiration from young infants who learn to talk before they learn to read or write. Here, the task is to discover a pseudo-text (a sub-word symbolic representation internal to the machine), from raw speech, without any labels, and to use these discovered units to learn higher order representations (lexical, syntactic, semantic) and achieve good scores in simple LM tasks.You will find a more complete description here: https://zerospeech.com/2021/ and a paper here: https://arxiv.org/abs/2011.11588
We would be happy to know whether you intend to participate in this challenge (this is for the calibration of our compute resources for the evaluation phase). If not it would be most useful to know what would be the major obstacle preventing you to participate (computer resources, time, etc). You can answer directly to this email: mailto:zerospeech2021@gmail.comThank youfor the organizersEmmanuel Dupoux--****************************************Emmanuel Dupoux
Directeur d'Etudes à l'Ecole des Hautes Etudes en Sciences SocialesEquipe Cognitive Machine LearningEHESS - ENS - CNRS - INRIA
29 rue d'Ulm, Paris 75005, France