Most state-of-the-art Deep Learning (DL) approaches for speaker recognition work on a short utterance level. Given the speech signal, these algorithms extract a sequence of speaker embeddings from short segments and those are averaged to obtain an utterance level speaker representation. In this work we propose the use of an attention mechanism to obtain a discriminative speaker embedding given non fixed length speech utterances. Our system is based on a Convolutional Neural Network (CNN) that encodes short-term speaker features from the spectrogram and a self multi-head attention model that maps these representations into a long-term speaker embedding. The attention model that we propose produces multiple alignments from different subsegments of the CNN encoded states over the sequence. Hence this mechanism works as a pooling layer which decides the most discriminative features over the sequence to obtain an utterance level representation. We have tested this approach for the verification task for the VoxCeleb1 dataset. The results show that self multi-head attention outperforms both temporal and statistical pooling methods with a 18% of relative EER. Obtained results show a 58% relative improvement in EER compared to i-vector+PLDA.

No abstract

Restricted Boltzmann Machines (RBMs) have shown success in different stages of speaker recognition systems. In this paper, we propose a novel framework to produce a vector-based representation for each speaker, which will be referred to as RBMvector. This new approach maps the speaker spectral features to a single fixed-dimensional vector carrying speaker-specific information. In this work, a global model, referred to as Universal RBM (URBM), is trained taking advantage of RBM unsupervised learning capabilities. Then, this URBM is adapted to the data of each speaker in the development, enrolment and evaluation datasets. The network connection weights of the adapted RBMs are further concatenated and subject to a whitening with dimension reduction stage to build the speaker vectors. The evaluation is performed on the core test condition of the NIST SRE 2006 database, and it is shown that RBM-vectors achieve 15% relative improvement in terms of EER compared to i-vectors using cosine scoring. The score fusion with i-vector attains more than 24% relative improvement. The interest of this result for score fusion yields on the fact that both vectors are produced in an unsupervised fashion and can be used instead of i-vector/PLDA approach, when no data label is available. Results obtained for RBM-vector/PLDA framework is comparable with the ones from i-vector/PLDA. Their score fusion achieves 14% relative improvement compared to i-vector/PLDA.

No abstract

In this paper, we propose to discriminatively model target and impostor spectral features using Deep Belief Networks (DBNs) for speaker recognition. In the feature level, the number of impostor samples is considerably large compared to previous works based on i-vectors. Therefore, those i-vector based impostor selection algorithms are not computationally practical. On the other hand, the number of samples for each target speaker is different from one speaker to another which makes the training process more difficult. In this work, we take advantage of DBN unsupervised learning to train a global model, which will be referred to as Universal DBN (UDBN). Then we adapt this UDBN to the data of each target speaker. The evaluation is performed on the core test condition of the NIST SRE 2006 database and it is shown that the proposed architecture achieves more than 8% relative improvement in comparison to the conventional Multilayer Perceptron (MLP).