ED Mathématiques et Informatique
Deep learning for modeling time-varying audio effects with control parameters
by Yann BOURDIN (Institut national de recherche en informatique et en automatique - Bordeaux - Sud-Ouest)
The defense will take place at 14h30 - Ada Lovelace Centre INRIA, université de Bordeaux, 200 avenue de la Vieille Tour, 33405 Talence
in front of the jury composed of
- Pierrick LEGRAND - Professeur des universités - Bordeaux INP - Directeur de these
- Joshua REISS - Professeur des universités - Queen Mary University of London - Rapporteur
- Juliette CHABASSIER - Chargée de recherche - Inria Bordeaux Sud Ouest - Examinateur
- Lauri JUVELA - Assistant professor - Aalto University - Examinateur
- Louis BIGO - Professeur des universités - Bordeaux INP, LaBRI, SCRIME - Examinateur
- Thomas HUEBER - Directeur de recherche - GIPSA-lab - Examinateur
The unique sound of iconic audio gear is often sought by musicians, sound engineers, and vintage enthusiasts. Digital modeling of these devices is an active research area, offering portability, flexibility, and lower costs, thus making them accessible to a wider audience and preserving them for the long-term. Yet, capturing their unique sonic traits is hard due to their nonlinear and time-varying behaviors. Traditionally, modeling audio effects uses white-box methods that build equations from electronic schematics. These methods need deep circuit knowledge, requiring heavy labor and computing overhead. Over the past decade, grey-box and black-box methods emerged, needing little to no circuit knowledge, relying instead on input-output data. Specifically, deep learning methods using neural networks have drawn great interest and are highly researched. However, building neural architectures that accurately emulate complex audio effects while keeping the causality and low latency needed for real-time processing remains a major challenge. This thesis explores deep learning for black-box modeling of audio effects, focusing on time-dependent and time-varying systems, and integrating user controls. We first address modeling time-dependent audio effects using dynamic range compression as our main case study. To efficiently capture long temporal dependencies, we propose the Series-Parallel Temporal Modulation (SPTMod) architecture. This convolutional-recurrent architecture features a low-rate modulation path and an audio path. A key challenge is properly initializing internal states, causing discrepancies between training and inference. We propose the State Prediction Network (SPN), that estimates initial states from past signal contexts. Further, we couple state prediction with truncated backpropagation through time, boosting both training time and inference accuracy. MUSHRA listening tests confirm our models reach high perceptual fidelity compared to the target effect. We then extend our framework to time-varying audio effects, where internal parameters change freely from the input signal, typically driven by a low-frequency oscillator. These effects are hard to model because supervised learning strictly enforces phase synchronization between the model and target, often stalling convergence. We adapt our architecture into the Series-Parallel Time-Varying Modulation (SPTVMod) network, adding depthwise dilated convolutions for better frequency processing, and a discriminator for adversarial training. To solve phase alignment issues, we develop a two-stage training strategy. The first phase uses a generative adversarial network to learn the modulation shape without needing strict phase synchronization. The second leverages the SPN for supervised fine-tuning, successfully syncing the model's internal oscillation to the target audio. We also explore modeling magnetic tape recorders, showing that oversampling within convolutional blocks effectively reduces aliasing artifacts induced by nonlinearities. Finally, we further explore integrating user controls, ensuring the models stay faithful across their entire control parameter space. Adding controls requires adapting the model via conditioning methods like input concatenation, additive conditioning, hypernetworks, or Feature-wise Linear Modulation (FiLM). Beyond architecture, the sampling strategy to populate the control space of the training dataset must be considered. We run a broad experiment comparing various sampling strategies and dataset sizes. Results show that best candidate sampling using an L1 or fractional norm beats uniform sampling for a given dataset budget. This allows reducing dataset size and measurement time while keeping high fidelity. Observing the learned compression curves and transient responses confirms the trained models smoothly interpolate between control parameter configurations without odd or unstable behaviors.
ED Droit
From the fragmentation of the world to the powers of politic: politic as a means of order.
by Lény ESCOBAR (INSTITUT DE RECHERCHE MONTESQUIEU)
The defense will take place at 14h00 - Salle des thèses 16 avenue Léon Duguit 33600 Pessac
in front of the jury composed of
- Patrick TROUDE-CHASTENET - Professeur émérite - Université de Bordeaux - Directeur de these
- Thierry GONTIER - Professeur des universités - Université Jean Moulin Lyon 3 - Rapporteur
- Thierry MéNISSIER - Professeur des universités - Université Grenoble Alpes IPHIG - Rapporteur
- Frédérique RUEDA - Professeure des universités - Université de Bordeaux - Examinateur
The thesis put forward could be summed up in a single simple proposition: politic is a means of ordering the world. Whilst this formulation may not be sufficient, it leaves no doubt as to the definitional perspective adopted or the subject of research: politics. Consequently, this work will argue that politics is an intermediate reality defined by its powers. Taken in this way, the thesis in question may seem obscure, which is not unrelated to the definitional perspective: in summarizing the work carried out, a large part of the discourse is absent. Whilst this work places the classic question ‘what is politics?' at its heart, it is equally shaped by the starting point from which this thesis emerges. Indeed, by employing the phrase ‘the tearing apart of the world', we begin from the premise that politics is merely one means among others of ordering the world. Without knowledge of, or control over, its origin and end, the political would be, strictly speaking, an intermediary present in this world. But whilst it is merely a means, it cannot be anything: it would be precisely a means, an intermediary, with its own reality. By supporting the genuine unity of politics, the aim is to examine the multiple and seemingly contradictory meanings attributed to it. At the same time, the aim is to demonstrate that the classic question ‘what is politics?' remains a topical and multifaceted issue which, even when defined in specific terms, can be approached in many different ways.This work therefore proposes to conceive of the political according to no chronology other than its own. This path is not without its difficulties, ranging from the possibility of conceiving of politics to that of defining this work. In this sense, in proposing this thesis, we cannot avoid the boundaries of the political, right down to the identification of this type of work. Furthermore, we shall seek to argue that an interest in borders proves fruitful when we examine the relationship between politics – specifically in the sense of political science – and the political.To this end, we propose to begin by focusing on myth. Why myth? Myth has a primary value of self-objection, stemming from the notion of the world's fragmentation, whilst we can find it in a diffuse form within political studies. By adopting a comparative approach, we shall seek to identify the meaning of myth from a political perspective; a meaning that will be termed ‘mythical' to define the scope of the study. The value of the mythical lies in offering a mode of expression with heuristic value: far from being mere fabrication, the mythical plays a part in teaching and persuasion. We thus arrive at the hypothesis that the mythical concerns both politics and political science. In order to address this particular hypothesis effectively, we shall draw on a modest genealogy of the political, seeking to trace its origin and its end. This approach, now widely adopted, stems from the idea that we can divide the political into its constituent moments. By examining various terms used to denote origin and end, such as initium or death, we shall identify the origin of the political in ‘the art of governing' and its end in its autonomy. However, it is only through a third phase, which aims to revisit origin and end, that we can argue that the political is an intermediate reality—that is, a produced reality that is neither purely intelligible nor purely sensible. This approach paves the way for a third section, which will focus on bringing together and summarizing the preceding sections by examining the concept of power. It is by dividing power into three political powers—maintaining, limiting and governing—that we shall be able to grasp and understand famous concepts of political science, such as those of the State or the Law, in a different light. It is with the help of this final framework that we shall succeed in upholding this simple thesis: politic is a means of ordering the world.
ED Sociétés, Politique, Santé Publique
Inquiry processes in organizational interventions: a challenge for collective reflexivity and the development of service activities
by Clara SIMILOWSKI (Laboratoire de l'Intégration du Matériau au Système)
The defense will take place at 14h00 - Amphithéâtre E Campus de la Victoire, 3ter place de la Victoire, 33000 Bordeaux
in front of the jury composed of
- Johann PETIT - Maître de conférences - Université de Bordeaux - Directeur de these
- Justine ARNOUD - Professeure des universités - Université Paris Est Créteil - CoDirecteur de these
- Fabien COUTAREL - Professeur des universités - Université Clermont Auvergne (UCA) - Rapporteur
- Nathalie RAULET-CROSET - Professeure des universités - IAE Paris Sorbonne Business School, Université Paris I Panthéon-Sorbonne - Rapporteur
- Vanina MOLLO - Maîtresse de conférences - Université de Toulouse - Examinateur
- Benoît JOURNE - Professeur des universités - Université de Nantes - Examinateur
Organizational consulting is one of those services that involve working not only for, but also with, a variety of actors. In consulting firms that promote the co-production of services—and in contrast to “expert” models—consultants strive to support client actors both in meeting their immediate needs and in fostering the development of skills over the longer term. This co-production approach makes the service more participatory, even democratic, but it is also more unsettling for consultants, particularly because the ends and means of the intervention must be redefined. The course and outcomes of interventions become uncertain (Bourgoin, 2013; Tran Van, 2010), and the plurality of actors and objectives is a source of instability (Flageul-Caroly, 2001; Petit, 2005). Intervention teams are thus confronted with so-called “indeterminate” situations (Dewey, 1938/1967) that challenge their established work practices (Arnoud, 2023; Lorino, 2013). Managing these situations becomes possible through collective reflexivity involving various actors within the client organization, as well as within the consulting firm itself. In the firm studied, organizational structures were designed to share complex situations with all actors within the firm. To examine consultants' collective reflexivity on and in their interventions, as well as their use of these organizational structures, we employ an analytical framework drawn from pragmatist philosophy: the inquiry process. Inquiry, mediated by dialogic activities (Bakhtin, 1984; Lorino et al., 2011), aims to transform an indeterminate situation into a more intelligible and actionable one, while producing shared reference points for action and collective learning. This thesis argues that consultants conduct inquiries to resolve ambiguous situations across multiple stages of intervention and with various actors who together form an “inquiry community.” The action research conducted through immersion in the firm allowed for the collection of data through observations, interviews, and discussions concerning the consultants' activities, their interventions, and the organizational structures designed to foster collective and dialogic reflexivity. The analyses of interventions are based on narrative accounts of inquiries conducted across three different interventions. Our results show that inquiry processes contribute to redesigning interventions. Modifications occur notably through a process of valuation (Dewey, 1939/2011), consisting in jointly redefining the ends and means with the actors of the inquiry community. Organizational structures can also contribute to consultants' ongoing inquiry by expanding the “public” to include the consulting firm's staff. However, these structures may also hinder ongoing inquiry and produce effects opposite to those intended, particularly when consultants express frustration or confusion. Our results further highlight the contribution of these inquiries to the consolidation of guidelines and know-how among the firm's professionals, which we refer to as a “community of practice.” Organizational structures thus serve as a channel for the transfer and development of skills among consultants. Furthermore, inquiries during interventions raise issues regarding service quality—helping consultants provide effective advice to client organizations—as well as health-related issues through the development of competencies. This empirical material enables a discussion of the organizational conditions necessary for conducting inquiries and, more broadly, for structuring collective reflexivity within organizations.