March 2026 – Laboratoire Informatique d’Avignon

Seminar – Matthew Wiesner – 26/03/2026

26 March 2026

Room 5 – 12h00 Title: Modeling Accent and Language at Scale Abstract: State-of-the-art LID models work reliably for ∼100 languages, but there are orders of magnitude more accents and dialects; annotating all of them is intractable. Furthermore, current LID models fail dramatically when applied to accented speech, and very few annotated accented data exist to support accented TTS. This talk explores these challenges and proposes a potential partial solution using massive amounts of diverse and widely available data collected from radio, with soft language labels in the form of geolocations. The talk then explores the link between robustness to accented speech and the capacity to more effectively model sequence-level data. Finally, these models and insights are shown to greatly improve LID on accented speech, and can be used to mine for accented speech to support scalable, controllable accented TTS.Bio: Matthew Wiesner est chercheur à l’Université Johns Hopkins et chercheur invité au Laboratoire Interdisciplinaire des Sciences du Numérique (LISN). Il a obtenu son doctorat et son master en génie électrique à Johns Hopkins en 2021 et 2016 sous la direction de Jan Trmal et de Sanjeev Khudanpur, ainsi que sa licence en génie électrique à l’Université McGill en 2013. Ses recherches Plus d'infos

PhD Defense – Manh Tuan NGUYEN – 25/03/2026

19 March 2026

Date: Wednesday, March 25th at 2:00 PM Place: amphitheater Blaise at CERI. The presentation will be held in English. The jury members: Title: Considering the inter-judge variability in the perceptual evaluation of speech and voice disorders and its integration into an automatic decision support system Abstract:Perceptual judgments are widely used in domains that lack clear objective criteria or reliable measurement methods, requiring reliance on human expert evaluation. However, such judgments are inherently subjective, often leading to a lack of agreement and, consequently, variability when multiple experts assess the same material. This variability, referred to as inter-rater variability, is typically addressed by aggregating scores or applying majority voting to produce a consensus decision. While effective for obtaining a final decision, this approach leaves the underlying causes of inter-rater variability largely unexplored.This thesis aims to explain inter-rater variability rather than treating consensus decisions as an absolute reference. We argue that such variability may arise from systematic differences between experts, particularly in terms of professional background, training, and the perceptual dimensions emphasized during assessment. By reducing individual judgments to a single consensus score, traditional approaches implicitly discard valuable information about expert reasoning and decision-making strategies. Understanding and explaining this variability is therefore essential.To address this objective, we propose a computational approach to model and interpret inter-rater variability. Leveraging the pattern-recognition capabilities of Plus d'infos

CORNET Seminar – Younes Ben Mazziane – 20/03/2026

16 March 2026

Date: 20/03/2026 à 11h30 Location: Salle C057 Title: Learning in Proportional Allocation Auctions Games Abstract: The Kelly or proportional allocation mechanism is a simple and efficient auction-based scheme that distributes an infinitely divisible resource proportionally to the agents’ bids. When agents are aware of the allocation rule, their interactions form a game, that has been extensively studied. This paper examines the less explored repeated Kelly game, focusing mainly on utilities that are logarithmic in the allocated resource fraction. We first derive this logarithmic form from fairness–throughput trade-offs in wireless network slicing, and then prove that the induced stage game admits a unique Nash equilibrium (NE). For the repeated play, we prove convergence to this NE under three behavioral models: (i) all agents use Online Gradient Descent (OGD), (ii) all agents use Dual Averaging with a quadratic regularizer (DAQ) (a variant of the Follow-the-Regularized leader algorithm), and (iii) all agents play myopic best responses (BR). Our convergence results hold even when agents use personalized learning rates in OGD and DAQ (e.g., tuned to optimize individual regret bounds), and they extend to a broader class of utilities that meet a certain sufficient condition. Finally, we complement our theoretical results with extensive simulations Plus d'infos

SLG Seminar – Tom Labiausse – 16/03/2026

12 March 2026

16 mars à 13h Tom présentera ses derniers travaux sur Hibiki-Zero, un modèle de traduction simultanée de la parole vers la parole. Les améliorations par rapport à ses travaux précédents, Hibiki, sont vraiment très intéressantes et s’appuient sur une technique d’apprentissage par renforcement (GRPO). Utilisée comme le propose Tom, cette technique évite d’avoir à préparer des données d’apprentissage parole-parole alignées au niveau mot comme dans la première mouture d’Hibiki. Vous êtes invités à assister à cette présentation qui pourrait vous donner des idées d’application à certains de vos travaux. Pour plus d’information, vous pouvez consulter cette page web très instructive et accessible (exemples, code, article) :