Theo Deschamps-Berger

Machine Learning Research Engineer at Diabolocom, specializing in speech representation learning and affective computing. I hold a Ph.D. from Paris-Saclay University, where my doctoral research investigated Transformer-based architectures—specifically the cross-modal alignment properties of Masked Language Modeling (MLM) for speech emotion recognition. My research interests include multimodal alignments between speech and text using raw information or neural audio codecs and optimizing training modeling (e.g., cumulative batches). I am also interested in developing transfer-learning guidelines between decoder and encoder models.

Machine Learning Research Engineer

Oct. 2024 - Present

Diabolocom Paris, France

Ph.D. in Computer Science

Oct. 2020 - Sept. 2024

Paris-Saclay University Orsay, France

Research at LISN (CNRS) on speech masked language models and transfer learning.
Funding: AI Chair, HUMAAINE Project

Master’s Research Intern

March - Sept. 2020

LISN (CNRS) Orsay, France

Social emotion recognition in call center conversations.

MSc in Artificial Intelligence

Sept. 2019 - Sept. 2020

Paris-Saclay University Orsay, France

Core: Deep Learning, NLP, Large-Scale Learning, Signal Processing.

Published in Langages, vol. 234, no. 2, 2024.

Les émotions‹ in the wild› des appelants d’un centre d’appels d’urgence: vers un système de détection des émotions dans la voix

Cet article explore la détection des émotions ‘in the wild’ dans les appels d’urgence.

PhD Thesis, Université Paris-Saclay, 2024.

Social Emotion Recognition with multimodal deep learning architecture in emergency call centers

This thesis presents a multimodal deep learning architecture for social emotion recognition in the context of emergency call centers.

Presented at ICMI 2023

A multi-task, multi-modal approach for predicting categorical and dimensional emotions

A joint multi-task and multi-modal approach for robust emotion prediction.

Presented at ICASSP 2023

Exploring attention mechanisms for multimodal emotion recognition in an emergency call center corpus

Investigation into attention-based models to integrate multiple modalities for emotion detection in real-life emergency scenarios.

Presented at ICMI 2023

Multiscale contextual learning for speech emotion recognition in emergency call center conversations

This paper explores multiscale contextual learning for improved speech emotion recognition in complex emergency call center dialogues.

Presented at ICMI 2022

Investigating Transformer Encoders and Fusion Strategies for Speech Emotion Recognition in Emergency Call Center Conversations

Analysis of various Transformer-based architectures and fusion techniques for emotion recognition in the noisy and emotional environment of emergency calls.

Presented at ACIIW 2021

Emotion Recognition in Emergency Call Centers: The challenge of real-life emotions

Discussing the unique challenges posed by real-life emotional extremes in emergency call settings.

Presented at ACII 2021

End-to-end speech emotion recognition: challenges of real-life emergency call centers data recordings

Addressing the technical and practical challenges of end-to-end emotion recognition using real-world acoustic data from emergency call centers.