Search

Recent advances in Deep Learning (DL) for speaker recognition have improved the performance but are constrained to the need of labels for the background data, which is difficult in prac- tice. In i-vector based speaker recognition, cosine (unsuper- vised) and PLDA (supervised) are the basic scoring techniques, with a big performance gap between the two. In this thesis we tried to fill this gap without using speaker labels in several ways. We applied Restricted Boltzmann Machine (RBM) vectors for the tasks of speaker clustering and tracking in TV broadcast shows. The experiments on AGORA database show that us- ing this approach we gain a relative improvement of 12% and 11% for speaker clustering and tracking tasks, respectively. We also applied DL techniques in order to increase the discrimina- tive power of i-vectors in speaker verification task, for which we have proposed the use of autoencoder in several ways, i.e., (1) as a pre-training for a Deep Neural Network (DNN), (2) as a near- est neighbor autoencoder for i-vectors, (3) as an average pooled nearest neighbor autoencoder. The experiments on VoxCeleb database show that we gain a relative improvement of 21%, 42% and 53%, using the three systems respectively. Finally we also proposed a self-supervised end-to-end speaker verification system. The architecture is based on a Convolutional Neural Network (CNN), trained as a siamese network with multiple branches. From the results we can see that our system shows comparable performance to a supervised baseline, This work was supported by the project PID2019-107579RBI00 / AEI / 10.13039/501100011033, Peer Reviewed, Postprint (published version)

In the last years, i-vectors followed by cosine or PLDA scoringtechniques were the state-of-the-art approach in speaker veri-fication. PLDA requires labeled background data, and thereexists a significant performance gap between the two scoringtechniques. In this work, we propose to reduce this gap by us-ing an autoencoder to transform i-vector into a new speaker vec-tor representation, which will be referred to as ae-vector. Theautoencoder will be trained to reconstruct neighbor i-vectors in-stead of the same training i-vectors, as usual. These neighbori-vectors will be selected in an unsupervised manner accordingto the highest cosine scores to the training i-vectors. The evalua-tion is performed on the speaker verification trials of VoxCeleb-1 database. The experiments show that our proposed ae-vectorsgain a relative improvement of 42% in terms of EER comparedto the conventional i-vectors using cosine scoring, which fillsthe performance gap between cosine and PLDA scoring tech-niques by 92%, but without using speaker labels, Peer Reviewed, Postprint (published version)

Most state-of-the-art Deep Learning (DL) approaches forspeaker recognition work on a short utterance level. Given thespeech signal, these algorithms extract a sequence of speakerembeddings from short segments and those are averaged to ob-tain an utterance level speaker representation. In this work wepropose the use of an attention mechanism to obtain a discrim-inative speaker embedding given non fixed length speech utter-ances. Our system is based on a Convolutional Neural Network(CNN) that encodes short-term speaker features from the spec-trogram and a self multi-head attention model that maps theserepresentations into a long-term speaker embedding. The atten-tion model that we propose produces multiple alignments fromdifferent subsegments of the CNN encoded states over the se-quence. Hence this mechanism works as a pooling layer whichdecides the most discriminative features over the sequence toobtain an utterance level representation. We have tested thisapproach for the verification task for the VoxCeleb1 dataset.The results show that self multi-head attention outperforms bothtemporal and statistical pooling methods with a18%of rela-tive EER. Obtained results show a58%relative improvementin EER compared to i-vector+PLDA, Peer Reviewed, Postprint (published version)

This work aims at creating expressive voices from audiobooks using semantic selection. First, for each utterance of the audiobook an acoustic feature vector is extracted, including iVectors built on MFCC and on F0 basis. Then, the transcription is projected into a semantic vector space. A seed utterance is projected to the semantic vector space and the N nearest neighbors are selected. The selection is then filtered by selecting only acoustically similar data. The proposed technique can be used to train emotional voices by using emotional keywords or phrases as seeds, obtaining training data semantically similar to the seed. It can also be used to read larger texts in an expressive manner, creating specific voices for each sentence. That later application is compared to a DNN predictor, which predicts acoustic features from semantic features. The selected data is used to adapt statistical speech synthesis models. The performance of the technique is analyzed objectively and in a perceptive experiment. In the first part of the experiment, subjects clearly show preference for particular expressive voices to synthesize semantically expressive utterances. In the second part, the proposed method is shown to achieve similar or better performance than the DNN based prediction. Copyright © 2016 ISCA., Peer Reviewed, Postprint (published version)