Search

Data augmentation is an effective method to increase the quantity of training data, which improves the model’s robustness and generalization ability. In this paper, we propose a generative adversarial network (GAN) based data augmentation approach for probabilistic linear discriminant analysis (PLDA), which is a standard back-end for state-of-the-art x-vector based speaker verification system. Instead of generating new spectral feature samples, a conditional Wasserstein GAN is adopted to directly generate x-vectors. Experiments are carried out on the standard NIST SRE 2016 evaluation dataset. Compared to manually adding noise, the GAN augmented PLDA achieves better performance and this performance can be further boosted when combined with manual augmented data. EER of 11.68% and 4.43% were obtained for Tagalog and Cantonese evaluation condition, respectively.

Probabilistic Linear Discriminant Analysis (PLDA) has exhibited good performance in face recognition and speaker verification. However, it is not widely used as a general-purpose classifier. The major limitation of PLDA lies in that, in the original formulation, the modeling part and the prediction part require the inversion of large matrices, whose sizes are proportional to the number of training vectors in a class. The original formulation of PLDA is not scalable if there are many training vectors, because the matrices will become too large to be inverted. In the literature, some scalable versions for the modeling part have been proposed. In this paper, we propose the scalable version for the prediction part, which completes the scalable version of PLDA. This makes PLDA able to handle a large number of training data, enabling PLDA to be used as a general-purpose classifier for different classification tasks. We then apply PLDA as the classifier to three different audio signal classification tasks, and compare its performance with Support Vector Machine (SVM), which is a widely used general-purpose classifier. Experimental results show that PLDA performs very well and can be even better than SVM, in terms of classification accuracy.

Conventional speaker diarization systems based on factor analysis mainly differ in i-vector scoring, such as the cosine scoring and the newly emerged probabilistic linear discriminant analysis (PLDA) scoring technique. However, during the clustering process, the accuracy of PLDA scoring decreases in short speech segments. The matter becomes even worse when the segments are with arbitrary duration. In this paper, we choose a modified PLDA model, called full posterior distribution PLDA (FP-PLDA) for clustering instead of the standard PLDA model (Std-PLDA). The new model exploits the intrinsic uncertainty of the i-vector extraction. The experiment shows that FP-PLDA has an especially effective performance in the short and variable duration speech segments. It relatively decreases the diarization error rate by around 41% for the cosine scoring system and 30.98% for the standard PLDA system.

For i-vector model, normalization approach is Probabilistic linear discriminant analysis and has a significant performance for verification of speaker. However it requires a huge development data which cost a lot in many cases. Unsupervised adaption method is a possible approach, which use unlabeled data to adapt PLDA scattering matrices to the target domain. In this paper, ‘local training’ approach is adapted to train the PLDA model. In this approach local labels discriminate the speaker within a single conversion and therefore easy to obtain as compare to normal ‘global labels’. Proposed approach can deliver better performance and improvement, particularly with limited globally-labeled data.

This paper compares the use of signal to noise ratio (SNR)-dependent and SNR-independent mixtures of probabilistic linear discriminant analysis (PLDA) versus conventional PLDA, under multi-noise and multi-SNR conditions for a small-set speaker verification system. Results indicate that conventional PLDA is more robust under multi-SNR conditions. The effect of the testing speech length is also examined and speech signals with a length of 5 seconds were found to achieve acceptable results.

One of the major reasons for the performance degradation of a speaker verification (SV) system in real-world conditions is its inability to spot speech regions due to the presence of noise. This work focuses on the role of voice activity detection (VAD) methods in alleviating such shortcomings. The experiments are conducted on the core-core task of the speakers in the wild (SITW) challenge. Two VAD approaches are explored in this work. One of them is the recently proposed self-adaptive VAD and the other is based on vowel-like region (VLR) detection. For evaluating the effectiveness of these approaches, the SV systems are developed using the i-vector framework in the front-end and probabilistic linear discriminant analysis (PLDA) in the back-end. The self-adaptive VAD based system shows better performance compared to the VLR based system in high SNR condition. Under degraded conditions, the VLR based method is relatively more robust compared to self-adaptive VAD. Exploiting these complementary features, significant improvements in the SV performances are noted with the fusion of scores of the two systems.

State-of-the-art speaker verification systems model speaker identity by mapping i-Vectors onto a probabilistic linear discriminant analysis (PLDA) space. Compared to other modeling approaches (such as cosine distance scoring), PLDA provides a more efficient mechanism to separate speaker information from other sources of undesired variabilities and offers superior speaker verification performance. Unfortunately, this efficiency is obtained at the cost of a required large corpus of labeled development data, which is too expensive/unrealistic in many cases. This study investigates a potential solution to resolve this challenge by effectively utilizing unlabeled development data with universal imposter clustering. The proposed method offers +21.9% and +34.6% relative gains versus the baseline system on two public available corpora, respectively. This significant improvement proves the effectiveness of the proposed method.

Speaker diarization via unsupervised i-vector clustering has gained popularity in recent years. In this approach, i-vectors are extracted from short clips of speech segmented from a larger multi-speaker conversation and organized into speaker clusters, typically according to their cosine score. In this paper, we propose a system that incorporates probabilistic linear discriminant analysis (PLDA) for i-vector scoring, a method already frequently utilized in speaker recognition tasks, and uses unsupervised calibration of the PLDA scores to determine the clustering stopping criterion. We also demonstrate that denser sampling in the i-vector space with overlapping temporal segments provides a gain in the diarization task. We test our system on the CALLHOME conversational telephone speech corpus, which includes multiple languages and a varying number of speakers, and we show that PLDA scoring outperforms the same system with cosine scoring, and that overlapping segments reduce diarization error rate (DER) as well.

Previous works have shown the benefits of empirical likelihood ratio (LR) kernels for i-vector/PLDA speaker verification. The method not only utilizes the multiple enrollment utterances of target speakers effectively, but also opens up opportunity for adopting sparse kernel machines for PLDA-based speaker verification systems. This paper proposes taking the advantages of the empirical LR kernels by incorporating them into relevance vector machines (RVMs). Results on NIST 2012 SRE demonstrate that the performance of RVM regression equipped with empirical LR kernels is slightly better than that of the support vector machines after performing utterance partitioning.

Probabilistic linear discriminant analysis (PLDA) has shown to be an effective model for disentangling speaker and channel variability in the i-vector space for text-independent speaker verification. The speaker and channel subspaces in the PLDA model are typically trained by optimizing the maximum likelihood (ML) criterion. PLDA assumes that i-vectors are normally distributed, which has shown to be violated in practice. This paper advocates the use of discriminative training, in which both target and non-target classes are taken into account to re-train the parameters. The efficacy of the proposed method is confirmed via experiments conducted on common condition 1 and 5 of the core task as specified in the Speaker Recognition Evaluations (SREs) 2010 conducted by the National Institute for Standards and Technology (NIST).

Intermodality face matching or Heterogeneous face recognition involves matching faces from different modalities such as infrared images, sketch images and low/high resolution visual images. This problem is further alleviated due to inherit problems in face recognition such as pose, expression, illumination, occlusion etc. Existing face recognition algorithms fail to address the existing feature gap exist between images of different modalities. To solve this problem, we propose a new method inspired from Probabilistic Linear Discriminant Analysis (PLDA). PLDA is a generative probabilistic method which models the face into signal and noise components. This method reports outstanding results when compared to other contemporary approaches. But PLDA is designed to apply the image data in only one modality. In this paper, its efficacy has been extended to more generic problem of handling faces captured in different modalities. Experiments conducted on HFB (VIS-NIR), Biosecure (Low-High or Webcam-Digitalcam) face databases validate its robustness and superiority over other methods.

Probabilistic Linear Discriminant Analysis (PLDA), used particularly in image and speech processing for face and speaker recognition, respectively, is a generative model requesting lots of data to be trained. In the paper several enhancements concerning the implementation of the estimation algorithm of PLDA are proposed providing substantial computational savings. At first, an inverse of a huge matrix is replaced by an inversion of two significantly smaller matrices. Subsequently, it is shown how to avoid the need to process the whole data set in each iteration of the estimation algorithm. Supplementary results are presented on NIST SRE 2008.

Facial expression is one of the main issues of face recognition in uncontrolled environments. In this paper, we apply the probabilistic linear discriminant analysis (PLDA) method to recognize faces across expressions. Several PLDA approaches are tested and cross-evaluated on the Cohn-Kanade and JAFFE databases. With less samples per gallery subject, high recognition rates comparable to previous works have been achieved indicating the robustness of the approaches. Among the approaches, the mixture of PLDAs has demonstrated better performances. The experimental results also indicate that facial regions around the cheeks, eyes, and eyebrows are more discriminative than regions around the mouth, jaw, chin, and nose.

Recently, i-vector extraction and Probabilistic Linear Discriminant Analysis (PLDA) have proven to provide state-of-the-art speaker verification performance. In this paper, the speaker verification score for a pair of i-vectors representing a trial is computed with a functional form derived from the successful PLDA generative model. In our case, however, parameters of this function are estimated based on a discriminative training criterion. We propose to use the objective function to directly address the task in speaker verification: discrimination between same-speaker and different-speaker trials. Compared with a baseline which uses a generatively trained PLDA model, discriminative training provides up to 40% relative improvement on the NIST SRE 2010 evaluation task.