Étude des techniques de deep learning en reconnaissance automatique du locuteur

Speaker Recognition (SRE), sometimes referred to as speaker biometrics, involves the identification, verification (authentication), classification and, by extension, segmentation, tracking and detection of speakers. It is a generic term used for any procedure involving the recognition of a person's identity based on their voice. In this context, deep learning has attracted much more interest from speech processing researchers, and it has recently been introduced in SRE. In most cases, deep learning models are adapted from automatic speech recognition (ASR) applications and applied to SRE, and have shown their ability to compete with state-of-the-art approaches. Nevertheless, the use of deep learning in SRE is still linked to ASR. On the other hand, deep learning models are now considered state-of-the-art in many areas of pattern recognition. In SRE, several architectures have been studied, such as deep neural networks (DNNs), deep belief networks (DBNs), restricted Boltzmann machines (RBMs), etc., while convolutional neural networks (CNNs) are the most widely used models in image processing.The objective of this thesis is to study deep learning models for the SRE domain. For this reason, we proposed a new way to use DBNs and DNNs in SRE, with the aim to extract the deep speaker features (DeepSF). Subsequently, we proposed a new usage of CNNs for the SRE problem. Although they are particularly designed for image processing problems, CNNs have recently been applied to SRE using spectrograms as input images. We believe that this approach is not optimal because it can lead to two cumulative errors in solving image processing and SRE problems. This is why we have developed a new method that allows the use of CNNs without using images.The results of the thesis represent an important finding to understand how deep learning models can be adapted to the problem of speaker recognition.