LISN · CNRS · Université Paris-Saclay
Ph.D. student at LISN (CNRS – Université Paris-Saclay), working on emotional adaptation and evaluation of speech-to-speech large language models for assistive robotics, elderly care and AI ethics.
ABOUT
I am a PhD student at the Interdisciplinary Laboratory of Digital Sciences (LISN – CNRS, Université Paris-Saclay) since October 2025, supervised by Laurence Devillers, within the ANR HUMA-AI-NE chair.
My research focuses on emotional adaptation and evaluation of speech-to-speech large language models for safer and more controllable use by vulnerable users, particularly in the context of assistive robotics and elderly care.
I am particularly interested in multimodal foundation models, self-supervised learning, emotion recognition, human-robot interaction, as well as ethical and societal implications of artificial intelligence systems.
I graduated from the Master of Science in Informatics (MOSIG) at Université Grenoble Alpes, where I developed a strong background in computer science, artificial intelligence and applied mathematics.
RESEARCH WORK
Generative Information Retrieval with Autoregressive LLMs
Study of autoregressive language models for document identifier generation and information retrieval.
Comparison of Multimodal Emotional Valence Classification Capabilities between Foundation Models and Speech-to-Speech Models
Evaluation of emotional recognition capabilities of audio, text and speech-to-speech models on multilingual spontaneous speech datasets.
Wizard of Oz Corpus for Elderly Users
Creation of a multimodal interaction corpus between elderly users and a social robot to study emotional AI in real-world contexts.
TEACHING
Tutorial Classes – Formal Languages
Language theory, automata and grammars.
Practical Sessions – Introduction to C++
Practical programming sessions and student supervision.
TECHNICAL SKILLS
Large Language Models
Adaptation of generative LLMs into feature extractors (Moshi/Mimi).
Foundation Models
Feature extraction using self-supervised foundation models: WavLM and XLM-RoBERTa.
Multimodal AI
Design of unified and extensible data pipelines for audio and text modalities.
Data Engineering
Indexing and normalization of heterogeneous academic datasets (Multi-source Dataset Alignment).
Transfer Learning
Layer Freezing, Parameter-Efficient Fine-Tuning and adaptation of foundation models.
Human-Robot Interaction
Creation of a voice and video corpus involving elderly users based on Wizard of Oz interactions.
CONTACT
romalves@lisn.fr · GitHub · LinkedIn
LISN – UMR 9015 · Building 507, rue du Belvédère · 91405 Orsay, France