An iterative framework for unsupervised learning in the PLDA based speaker verification

Recent work reported on fast speaker search over large speech data corpora has focused on using locality sensitive hashing (LSH) search with hashing functions approximating i-vector based cosine distances (CosDist) for model comparisons. Because of the superior performance of probabilistic linear discriminant analysis (PLDA) model reported on speaker identification (SID) in recent years, in this paper we focus on using PLDA for fast speaker search. It is challenging to approximate PLDA well with simple hashing functions, resulting in difficulty to combine it with LSH search. As an alternative, we adopt a clustering-based pruning strategy to speed up PLDA search. Our results show the strategy can significantly speed up search with minimal performance loss. Another focus of this work is on PLDA model adaptation to mismatched conditions under which the fast search runs. The technique we adopt to adapt the PLDA model is based on the LDA adaptation method reported in [1], primarily adapting the LDA transform. Our results show this adaptation improves PLDA performance significantly (over 25% relative) on data collected in different conditions. Our speed-up experiments running with adapted LDA show that gains from the adapted PLDA are retained after the speed-up.

The Gaussian probabilistic linear discriminant anal-ysis (PLDA) model assumes Gaussian distributed priors for the latent variables that represent the speaker and channel factors. Assuming that each training i-vector belongs to a different speaker, as is usually done in i-vector extraction, i-vectors generated by a PLDA model can be considered independent and identically distributed with Gaussian distribution. Thus, we have recently proposed to transform the development i-vectors so that their distribution becomes more Gaussian-like. This is obtained by means of a sequence of affine and nonlinear transformations whose parameters are trained by maximum likelihood (ML) estimation on the development set. The evaluation i-vectors are then subject to the same transformation. Although the i-vector “gaussianization” has shown to be effective, since the i-vectors extracted from segments of the same speaker are not independent, the original assumption is not satisfactory. In this work, we show that the model can be improved by properly exploiting the information about the speaker labels, which was ignored in the previous model. In particular, a more effective PLDA model can be obtained by jointly estimating the PLDA parameters and the parameters of the nonlinear transformation of the i-vectors. In other words, while the goal of the previous approach was to “gaussianize” the training i-vectors distribution, the objective of this work is to embed the estimation of the nonlinear i-vector transformation in the PLDA model estimation. We will thus refer to this model as the nonlinear PLDA model. We show that this new approach provides significant gain with respect to PLDA, and a small, yet consistent, improvement with respect to our former i-vector “gaussianization” approach, without further additional costs.

In the i‐vector/probabilistic linear discriminant analysis (PLDA) technique, the PLDA backend classifier is modelled on i‐vectors. PLDA defines an i‐vector subspace that compensates the unwanted variability and helps to discriminate among speaker‐phrase pairs. The channel or session variability manifested in i‐vectors are known to be nonlinear in nature. PLDA training, however, assumes the variability to be linearly separable, thereby causing loss of important discriminating information. Besides, the i‐vector estimation, itself, is known to be poor in case of short utterances. This paper attempts to address these issues using a simple hierarchy‐based system. A modified fuzzy‐clustering technique is employed to divide the feature space into more characteristic feature subspaces using vocal source features. Thereafter, a separate i‐vector/PLDA model is trained for each of the subspaces. The sparser alignment owing to subspace‐specific universal background model and the relatively reduced dimensions of variability in individual subspaces help to train more effective i‐vector/PLDA models. Also, vocal source features are complementary to mel frequency cepstral coefficients, which are transformed into i‐vectors using mixture model technique. As a consequence, vocal source features and i‐vectors tend to have complementary information. Thus using vocal source features for classification in a hierarchy tree may help to differentiate some of the speaker‐phrase classes, which otherwise are not easily discriminable based on i‐vectors. The proposed technique has been validated on Part 1 of RSR2015 database, and it shows a relative equal error rate reduction of up to 37.41% with respect to the baseline i‐vector/PLDA system.

In this paper, we advocate the use of the uncompressed form of i-vector and depend on subspace modeling using probabilistic linear discriminant analysis (PLDA) in handling the speaker and session (or channel) variability. An i-vector is a low-dimensional vector containing both speaker and channel information acquired from a speech segment. When PLDA is used on an i-vector, dimension reduction is performed twice: first in the i-vector extraction process and second in the PLDA model. Keeping the full dimensionality of the i-vector in the i-supervector space for PLDA modeling and scoring would avoid unnecessary loss of information. We refer to the uncompressed i-vector as the i-supervector. The drawback in using the i-supervector with PLDA is the inversion of large matrices in the estimation of the full posterior distribution, which we show can be solved rather efficiently by portioning large matrices into smaller blocks. We also introduce the Gaussianized rank-norm, as an alternative to whitening, for feature normalization prior to PLDA modeling. We found that the i-supervector performs better during normalization. A better performance is obtained by combining the i-supervector and i-vector at the score level. Furthermore, we also analyze the computational complexity of the i-supervector system, compared with that of the i-vector, at four different stages of loading matrix estimation, posterior extraction, PLDA modeling, and PLDA scoring.

This study proposes an i-Vector based approach to gender identification. Gender-labeled utterances from the Fisher English (FE) corpus are used to formulate an i-Vector extraction framework, and a Probabilistic Linear Discriminant Analysis (PLDA) back-end is employed to compute the scores for gender identification. A novel duration mismatch compensation strategy is also presented that offers very little degradation in identification accuracy even with a large reduction in the duration of the test-segment. The proposed method is shown to consistently outperform a GMM-UBM based gender-identification scheme on several test-sets created from a held-out portion of the FE corpus, and is able to achieve an identification accuracy of up to 97.63%. On the severely distorted and multilingual DARPA-RATS (Robust Automatic Transcription of Speech) corpora, the proposed approach achieves an identification accuracy of 76.48% using only the FE data in training. Next, a novel unsupervised domain adaptation strategy is also presented that utilizes only unlabeled RATS data to adapt the out-of-domain PLDA parameters derived from the FE training data. The strategy is able to offer a 6.8% relative improvement in identification accuracy, and a 14.75% relative reduction in Equal Error Rate (EER) compared to using the out-of-domain PLDA model on the RATS test-utterances. These improvements are significant since: 1) RATS test-utterances are severely distorted, 2) No labeled data of any kind is used for 4 of the 5 languages present in the test-utterances.

This paper proposes a method for Bayesian estimation of probabilistic linear discriminant analysis (PLDA) when training labels are noisy. Label errors can be expected during e.g. large or distributed data collections, or for crowd-sourced data labeling. By interpreting true labels as latent random variables, the observed labels are modeled as outputs of a discrete memoryless channel, and the maximum a posteriori (MAP) estimate of the PLDA model is derived via Variational Bayes. The proposed framework can be used for PLDA estimation, PLDA domain adaptation, or to infer the reliability of a PLDA training list. Although presented as a general method, the paper discusses specific applications for speaker verification. When applied to the Speakers in the Wild (SITW) Task, the proposed method achieves graceful performance degradation when label errors are introduced into the training or domain adaptation lists. When applied to the NIST 2018 Speaker Recognition Evaluation (SRE18) Task, which includes adaptation data with noisy speaker labels, the proposed technique provides performance improvements relative to unsupervised domain adaptation.

The paper describes a multi-target speaker detection and identification system based on a fusion of probabilistic linear discriminant analysis (PLDA) and deep neural network (DNN). PLDA is the state-of-the-art approach used in speaker recognition, thus we selected it as our baseline. We tried to develop a DNN based approach, that would be more accurate than the baseline, but only better discrimination between blacklist and background speakers was achieved. The fusion of PLDA and DNN improved performance of the baseline system.

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.

Probabilistic linear discriminant analysis (PLDA) is widely described as an effective model for text-independent speaker verification in the i-vector space. The PLDA scoring function is typically formulated as the likelihood ratio between the speaker-adapted and the universal PLDAs. In this case, the adaptation of PLDA was performed through the speaker factors. In this paper, we show that the channel factors of the PLDA could be equivalently exploited to deal with the multi-source conditions. In speaker verification, with the proposed method, a PLDAmodel trained on conversational telephone speech could be adequately adapted for interview-style microphone recordings. Experimental results on NIST SRE'08 and SRE'10 datasets confirm that the proposed method is effective, especially for the case whereby enrollment and test utterances were captured from different sources.

Probabilistic linear discriminant analysis (PLDA) is the leading method for computing scores in speaker recognition systems. The method models the vectors representing each audio sample as a sum of three terms: one that depends on the speaker identity, one that models the within-speaker variability, and one that models any remaining variability. The last two terms are assumed to be independent across samples. We recently proposed an extension of the PLDA method, which we termed Joint PLDA (JPLDA), where the second term is considered dependent on the type of nuisance condition present in the data (e.g., the language or channel). The proposed method led to significant gains for multilanguage speaker recognition when taking language as the nuisance condition. In this paper, we present a generalization of this approach that allows for multiple nuisance terms. We show results using language and several nuisance conditions describing the acoustic characteristics of the sample and demonstrate that jointly including all these factors in the model leads to better results than including only language or acoustic condition factors. Overall, we obtain relative improvements in detection cost function between 5% and 47% for various systems and test conditions with respect to standard PLDA approaches. © 2018 International Speech Communication Association. All rights reserved.

In the biometrics community, face and speaker recognition are mature fields in which several systems have been proposed over the past twenty years. While existing systems perform well under controlled recording conditions, mismatch caused by the use of different sensors or a lack of cooperation from the subject still significantly affects performance, especially in challenging scenarios such as in forensics. Furthermore, existing methods suffer from scalability issues, which prevents them from taking advantage of increasingly large amounts of training data. This is otherwise a promising approach to improve accuracy in such challenging scenarios. In this thesis we address these problems of mismatch and complexity by developing scalable probabilistic models that we apply to face, speaker and bimodal recognition. Our contributions are four-fold. First, we propose a unified framework for session variability modeling techniques based on Gaussian mixture models (GMM), that encompasses inter-session variability (ISV) modeling, joint factor analysis (JFA) and total variability (TV) modeling. Second, we propose a novel exact and scalable formulation of probabilistic linear discriminant analysis (PLDA), which is a probabilistic and generative framework that models between-class and within-class variations. This formulation solves a major scalability issue, by improving both the time complexity of the training procedure from cubic to linear with respect to the number of samples per class, and the complexity of the scoring procedure. Furthermore, the implementations of all the proposed techniques are integrated into a novel collaborative open source software library called Bob 1 that enforces fair evaluations and encourages reproducible research. Fourth and finally, large-scale experiments are conducted with all of the above machine learning algorithms on several databases such as FRGC for face recognition, NIST SRE12 for speaker recognition and MOBIO for bimodal recognition, showing competitive performances.

Probabilistic linear discriminant analysis (PLDA) or cosine similarity have been widely used in traditional speaker verification systems as back-end techniques to measure pairwise similarities. To make better use of multiple enrollment utterances, we propose a novel attention back-end model, which can be used for both text-independent (TI) and text-dependent (TD) speaker verification, and employ scaled-dot self-attention and feed-forward self-attention networks as architectures that learn the intra-relationships of the enrollment utterances. In order to verify the proposed attention back-end, we conduct a series of experiments on CNCeleb and VoxCeleb datasets by combining it with several sate-of-the-art speaker encoders including TDNN and ResNet. Experimental results using multiple enrollment utterances on CNCeleb show that the proposed attention back-end model leads to lower EER and minDCF score than the PLDA and cosine similarity counterparts for each speaker encoder and an experiment on VoxCeleb indicate that our model can be used even for single enrollment case.

This paper analyses the short utterance probabilistic linear discriminant analysis (PLDA) speaker verification with utterance partitioning and short utterance variance (SUV) modelling approaches. Experimental studies have found that instead of using single long-utterance as enrolment data, if long enrolled utterance is partitioned into multiple short utterances and average of short utterance i-vectors is used as enrolled data, that improves the Gaussian PLDA (GPLDA) speaker verification. This is because short utterance i-vectors have speaker, session and utterance variations, and utterance-partitioning approach compensates the utterance variation. Subsequently, SUV-PLDA is also studied with utterance partitioning approach, and utterance partitioning-based SUV-GPLDA system shows relative improvement of 9% and 16% in EER for NIST 2008 and NIST 2010 truncated 10sec-10sec evaluation condition as utterance partitioning approach compensates the utterance variation and SUV modelling approach compensates the mismatch between full-length development data and short-length evaluation data.

This paper analyses the probabilistic linear discriminant analysis (PLDA) speaker verification approach with limited development data. This paper investigates the use of the median as the central tendency of a speaker's i-vector representation, and the effectiveness of weighted discriminative techniques on the performance of state-of-the-art length-normalised Gaussian PLDA (GPLDA) speaker verification systems. The analysis within shows that the median (using a median fisher discriminator (MFD)) provides a better representation of a speaker when the number of representative i-vectors available during development is reduced, and that further, usage of the pair-wise weighting approach in weighted LDA and weighted MFD provides further improvement in limited development conditions. Best performance is obtained using a weighted MFD approach, which shows over 10% relative improvement in EER over the baseline GPLDA system on mismatched and interview-interview conditions.

In this work, we assess the impact of vocal effort on discrimination and calibration performance of a state-of-the-art speaker recognition system. We analyze three levels of vocal effort (low, normal, and high) from the SRI-FRTIV corpus. We use a deep neural network (DNN) speaker embeddings system with probabilistic linear discriminant analysis (PLDA) and find that vocal effort variation significantly degrades system performance. We apply both mixture PLDA (mix-PLDA) and trial-based calibration with condition PLDA similarity (TBC-CPLDA) to improve system robustness. Our proposed approaches resulted in 18% and 33% relative improvement in discrimination and calibration performance respectively on the SRI-FRTIV corpus.