Deep Learning for i-Vector Speaker and Language Recognition

Over the last few years, i-vectors have been the state-of-the-art technique in speaker and language recognition. Recent advances in Deep Learning (DL) technology have improved the quality of i-vectors but the DL techniques in use are computationally expensive and need speaker or/and phonetic labels for the background data, which are not easily accessible in practice. On the other hand, the lack of speaker-labeled background data makes a big performance gap, in speaker recognition, between two well-known cosine and Probabilistic Linear Discriminant Analysis (PLDA) i-vector scoring techniques. It has recently been a challenge how to fill this gap without speaker labels, which are expensive in practice. Although some unsupervised clustering techniques are proposed to estimate the speaker labels, they cannot accurately estimate the labels. This thesis tries to solve the problems above by using the DL technology in different ways, without any need of speaker or phonetic labels. In order to fill the performance gap between cosine and PLDA scoring given unlabeled background data, we have proposed an impostor selection algorithm and a universal model adaptation process in a hybrid system based on Deep Belief Networks (DBNs) and Deep Neural Networks (DNNs) to discriminatively model each target speaker. In order to have more insight into the behavior of DL techniques in both single and multi-session speaker enrollment tasks, some experiments have been carried out in both scenarios. Experiments on the National Institute of Standard and Technology (NIST) 2014 i-vector challenge show that 46% of this performance gap, in terms of minDCF, is filled by the proposed DL-based system. Furthermore, the score combination of the proposed DL-based system and PLDA with estimated labels covers 79% of this gap. In the second line of the research, we have developed an efficient alternative vector representation of speech by keeping the computational cost as low as possible and avoiding phonetic labels, which are not always accessible. The proposed vectors will be based on both Gaussian Mixture Models (GMMs) and Restricted Boltzmann Machines (RBMs) and will be referred to as GMM-RBM vectors. The role of RBM is to learn the total speaker and session variability among background GMM supervectors. This RBM, which will be referred to as Universal RBM (URBM will then be used to transform unseen supervectors to the proposed low dimensional vectors. The use of different activation functions for training the URBM and different transformation functions for extracting the proposed vectors are investigated. At the end, a variant of Rectified Linear Unit (ReLU) which is referred to as Variable ReLU (VReLU) is proposed. Experiments on the core test condition 5 of the NIST Speaker Recognition Evaluation (SRE) 2010 show that comparable results with conventional i-vectors are achieved with a clearly lower computational load in the vector extraction process. Finally, for the Language Identification (LID) application, we have proposed a DNN architecture to model effectively the i-vector space of four languages, English, Spanish, German, and Finnish, in the car environment. Both raw i-vectors and session variability compensated i-vectors are evaluated as input vectors to DNN. The performance of the proposed DNN architecture is compared with both conventional GMM-UBM and i-vector/Linear Discriminant Analysis (LDA) systems considering the effect of the duration of signals. It is shown that the signals with a duration between 2 and 3 sec meet the accuracy and speed requirements of this application, in which the proposed DNN architecture outperforms GMM-UBM and i-vector/LDA systems by 37% and 28%, respectively.