[PhD defence] 24/10/2025 - Lucas Maison: "Robustness of neural models for automatic speech processing" (UPR LIA)

Lucas Maison will publicly defend his thesis on Friday 24 October 2025, entitled "Robustness of neural models for automatic speech processing"..

Laboratory

UPR 4128 LIA - Avignon Computing Laboratory

Composition of the jury

Mr Yannick ESTEVE	Avignon University	Thesis supervisor
Mr Benjamin LECOUTEUX	LIG/GETALP	Rapporteur
Ms Irina ILLINA	LORIA-INRIA	Rapporteur
Mr François CAPMAN	Thales SIX	Thesis co-supervisor
Mr Jean-François BONASTRE	Inria Defence & Security	Examiner
Ms Marcely ZANON BOITO	Naver Labs	Examiner

Summary

Automatic speech recognition has become a popular tool with many applications; it also serves as an intermediate step for other speech-related tasks, such as understanding spoken language or speech synthesis. In automatic speech recognition, the speech signal is first emitted by the speaker, transmitted through the environment, before being captured by a recording device and processed by a machine learning model. However, each of these stages can be a source of variability and lead to transcription errors, which affects the robustness of the system. In this thesis, we study various factors influencing the processing of speech by machines. More specifically, we focus on models pre-trained in French and refined for speech recognition. We begin by presenting our work on accent robustness. Through a number of experiments, we evaluate the resilience of the model to variations in accents and explore different ways of bridging the gaps between them. In particular, we examine the impact of the proportions of accented voices in the training set. In addition, we present CEREALES, a new dataset for Québécois French. In addition to accents, we are also interested in the impact of demographic variables on speech recognition performance. Using the Common Voice corpus, we highlight the model's biases and attempt to reduce them by using deliberately biased training sets. Finally, the last chapter explores the issue of acoustic robustness using keyword recognition models: we show how ID and OOD performance are correlated and study how training data or different pre-processing influence robustness.

Keywords : robustness, speech recognition, self-supervised learning, machine learning, speech