In a society where a considerable number of complex systems and connected objects are emerging, the need to simplify Human-Machine interactions has become an important issue for both the scientific community and companies. In this context, speech being considered as an obvious means of communication, the definition of an Automatic Speech Recognition (ASR) system meeting different criteria of performance, robustness and speed is crucial. It must also be scalable and easily adaptable to the evolution of a language, of a specific business context or of the connection with other software bricks of the communication chain (e.g., the modules for detecting user intentions or for understanding a textual statement). In this context, the main objective of this thesis is the development of an automatic speech recognition system for French and its linking with a natural language understanding system (NLC) proposed by the company Airudit, which is the sponsor of this thesis with the Laboratoire Bordelais de Recherche en Informatique (LaBRI). In order to answer the first objective, we propose a comparative study of the main existing NLP approaches for French today. Particular attention is given here to the type of approach (traditional PAR or end-to-end PAR), the definition of the optimal architecture as well as the type of output units (characters, sub-words or words). This study is then extended by a comparison of the errors formulated by the different systems produced for interpretation by a natural language understanding system. At the same time, I present my contributions to the ESPnet project which provides the scientific community with tools for speech processing. Still in the perspective of building an extit{optimal} PAR system for French, a particular attention has been given to the proposal of inference training techniques for one of the most performing systems during our first study: the RNN-Transducer. Concerning the second objective aiming at studying and improving the linkage with a comprehension system, we also propose a new approach that has been patented. This approach aims at restoring observed mechanisms of natural communication (e.g., a priori knowledge of the context, inference of meaning despite lack of information or presence of errors) and also proposes to solve problems inherent in the association of a traditional PAR system with a CLN system, created and optimized independently.