Probabilistic Approach Using Joint Long and Short Session i-Vectors Modeling to Deal with Short Utterances for Speaker Recognition

Kheder, Waad Ben; Matrouf, Driss; Ajili, Moez; Bonastre, Jean-François

doi:10.21437/interspeech.2016-1302

“…Previous work (Kheder et al, 2016;Guo et al, 2017b) highlights the importance of duration matching in PLDA model training. For instance when the PLDA is trained using long utterances and evaluated on short utterances, there is degradation in speaker verification performance compared to PLDA trained using matched-length short utterances.…”

Section: I-vector Mapping Resultsmentioning

confidence: 98%

“…A few recent papers have focused on i-vector mapping, which maps the short utterance i-vector to its long version. In Kheder et al (2016Kheder et al ( , 2018, the authors proposed a probabilistic approach, in which a GMM-based joint model between long and short utterance i-vectors was trained, and a minimum mean square error (MMSE) estimator was applied to transform a short i-vector to its long version. Since the GMM-based mapping function is actually a weighted sum of linear functions, our previous research (Guo et al, 2017b) demonstrates that a proposed non-linear mapping using convolutional neural networks (CNNs) outperforms the GMM-based linear mapping methods across different conditions.…”

Section: Introductionmentioning

confidence: 99%

Deep neural network based i-vector mapping for speaker verification using short utterances

Guo

¹

,

Xu

²

,

Qian

³

et al. 2018

Speech Communication

View full text Add to dashboard Cite

Text-independent speaker recognition using short utterances is a highly challenging task due to the large variation and content mismatch between short utterances. I-vector and probabilistic linear discriminant analysis (PLDA) based systems have become the standard in speaker verification applications, but they are less effective with short utterances. In this paper, we first compare two state-of-the-art universal background model (UBM) training methods for i-vector modeling using full-length and short utterance evaluation tasks. The two methods are Gaussian mixture model (GMM) based (denoted I-vector GMM) and deep neural network (DNN) based (denoted as I-vector DNN) methods. The results indicate that the I-vector DNN system outperforms the I-vector GMM system under various durations (from full length to 5 s). However, the performances of both systems degrade significantly as the duration of the utterances decreases. To address this issue, we propose two novel nonlinear mapping methods which train DNN models to map the i-vectors extracted from short utterances to their corresponding long-utterance i-vectors. The mapped i-vector can restore missing information and reduce the variance of the original short-utterance i-vectors. The proposed methods both model the joint representation of short and long utterance i-vectors: the first method trains an autoencoder first using concatenated short and long utterance i-vectors and then uses the pre-trained weights to initialize a supervised regression model from the short to long version; the second method jointly trains the supervised regression model with an autoencoder reconstructing the short utterance i-vector itself. Experimental results using the NIST SRE 2010 dataset show that both methods provide significant improvement and result in a 24.51% relative improvement in Equal Error Rates (EERs) from a baseline system. In order to learn a better joint representation, we further investigate the effect of a deep encoder with residual blocks, and the results indicate that the residual network can further improve the EERs of a baseline system by up to 26.47%. Moreover, in order to improve the short i-vector mapping to its long version, an additional vector, which represents the average value of phoneme posteriors across frames, is also added to the input, and results in a 28.43% improvement. When further testing the best-validated models of SRE10 on the Speaker In The Wild (SITW) dataset, the methods result in a 23.12% improvement on arbitrary-duration (1-5 s) short-utterance conditions.

show abstract

“…In the following two subsections we describe the GMM-based mapping algorithm as in [10] and our proposed CNN-based mapping algorithm in details.…”

Section: I-vector Mapping Between Short and Long Sessionsmentioning

confidence: 99%

“…(4), we can observe that the GMM-based mapping is actually a weighted sum of linear functions, and the weights are the conditional probabilities of each Gaussian component given test utterance x0. Even though the GMM-based joint modeling method gives significant improvement for the mismatched condition between short and long session i-vectors [10], there are still some shortcomings of this method. Learning a mapping from short session i-vector to its long version, is a very complex and nonlinear transform.…”

Section: Gmm-based Joint Probability Modelmentioning

confidence: 99%

“…In this paper, motivated by the good properties of long utterance i-vectors, we want to learn a mapping from the short utterance i-vector to a corresponding long utterance version, such that the improved version of i-vectors can be estimated from unseen short utterance i-vectors. To learn such mapping, the authors in [10] proposed a probabilistic approach, in which a GMM-based joint model between long and short utterance ivectors was trained, and a MMSE (minimum mean square error) estimator was applied to transform a short i-vector to its long version. Here, we propose a novel semi-supervised mapping algorithm which uses CNNs to extract useful information from short utterance i-vectors and maps it to a corresponding long version.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

CNN-Based Joint Mapping of Short and Long Utterance i-Vectors for Speaker Verification Using Short Utterances

Guo

¹

,

Nookala

²

,

Alwan

³

2017

View full text Add to dashboard Cite

Text-independent speaker recognition using short utterances is a highly challenging task due to the large variation and content mismatch between short utterances. I-vector and probabilistic linear discriminant analysis (PLDA) based systems have become the standard in speaker verification applications, but they are less effective with short utterances. To address this issue, we propose a novel method, which trains a convolutional neural network (CNN) model to map the i-vectors extracted from short utterances to the corresponding long-utterance i-vectors. In order to simultaneously learn the representation of the original short-utterance i-vectors and fit the target long-version ivectors, we jointly train a supervised-regression model with an autoencoder using CNNs. The trained CNN model is then used to generate the mapped version of short-utterance i-vectors in the evaluation stage. We compare our proposed CNNbased joint mapping method with a GMM-based joint modeling method under matched and mismatched PLDA training conditions. Experimental results using the NIST SRE 2008 dataset show that the proposed technique achieves up to 30% relative improvement under duration mismatched PLDA-training conditions and outperforms the GMM-based method. The improved systems also perform better compared with the matched-length PLDA training condition using short utterances.

show abstract

Duration compensation of i‐vectors for short duration speaker verification

Ma¹,

Sethu²,

et al. 2017

View full text Add to dashboard Cite

The standard i-vector/Gaussian probabilistic linear discriminant analysis (G-PLDA) system does not compensate for duration mismatch, which is a significant confounding factor in short duration speaker verification. A novel duration compensation technique to normalise the distribution mismatch caused by duration variation in the i-vector space is proposed. The proposed technique involves the use of two factor analysers that are tied together to share latent variables for a given speaker as the underlying generative model of the i-vector space. This leads to a transform which maps the original i-vectors onto a latent subspace that is expected to be duration invariant. The proposed method has the advantages that it normalises distribution mismatch while taking into consideration both inter-and intra-speaker variability. Experiments conducted on NIST SRE 2010 database shows that the proposed method leads to 18.54, 15.48 and 8.77% relative improvements when tested on utterances of 10, 5 and 3 s durations, respectively, compared with the best results obtained by either standard i-vector/G-PLDA or the previously proposed twin model G-PLDA.

show abstract

Probabilistic Approach Using Joint Long and Short Session i-Vectors Modeling to Deal with Short Utterances for Speaker Recognition

Cited by 9 publications

References 14 publications

Deep neural network based i-vector mapping for speaker verification using short utterances

Deep neural network based i-vector mapping for speaker verification using short utterances

CNN-Based Joint Mapping of Short and Long Utterance i-Vectors for Speaker Verification Using Short Utterances

Duration compensation of i‐vectors for short duration speaker verification

Contact Info

Product

Resources

About