Gaussian Probabilistic Linear Discriminant Analysis Research Articles

Generalized Domain Adaptation Framework for Parametric Back-End in Speaker Recognition

State-of-the-art speaker recognition systems comprise a speaker embedding front-end followed by a probabilistic linear discriminant analysis (PLDA) back-end. The effectiveness of these components relies on the availability of a large amount of labeled training data. In practice, it is common for domains (e.g., language, channel, demographic) in which a system is deployed to differ from that in which a system has been trained. To close the resulting gap, domain adaptation is often essential for PLDA models. Among two of its variants are Heavy-tailed PLDA (HT-PLDA) and Gaussian PLDA (G-PLDA). Though the former better fits real feature spaces than does the latter, its popularity has been severely limited by its computational complexity and, especially, by the difficulty, it presents in domain adaptation, which results from its non-Gaussian property. Various domain adaptation methods have been proposed for G-PLDA. This paper proposes a generalized framework for domain adaptation that can be applied to both of the above variants of PLDA for speaker recognition. It not only includes several existing supervised and unsupervised domain adaptation methods but also makes possible more flexible usage of available data in different domains. In particular, we introduce here two new techniques: (1) correlation-alignment in the model level, and (2) covariance regularization. To the best of our knowledge, this is the first proposed application of such techniques for domain adaptation w.r.t. HT-PLDA. The efficacy of the proposed techniques has been experimentally validated on NIST 2016, 2018, and 2019 Speaker Recognition Evaluation (SRE’16, SRE’18 and SRE’19) datasets.

A Robust Radio-Frequency Fingerprint Extraction Scheme for Practical Device Recognition

Radio-frequency fingerprinting (RFF) exploiting hardware characteristics has been employed for device recognition to enhance the overall security. However, the performance unreliability in long-term experiments, channel fading interference, and unauthorized devices verification are three open problems that restrict the development of RFF recognition. To address these issues, a robust RFF extraction scheme based on three corresponding algorithms is studied. For the first problem, a long-term stacking of repetitive symbols (LSRSs) algorithm is proposed to reduce the acquired signal variance, which contributes to the identification accuracy and long-term stability. For the second issue, we propose an artificial noise adding (ANA) algorithm to enhance the recognition robustness through regularization and channel adaptation. For the third issue, a verification algorithm based on the generative Gaussian probabilistic linear discriminant analysis (GPLDA) model is developed to handle unauthorized devices. Our robust RFF extraction scheme is verified in the experiments with 54 CC2530 ZigBee devices. It enables reliable node identification with the accuracy of 99.50% in the short rang line-of-sight (SLOS) scenarios for signals collected over 18 months, and 95.52% in the extensive multipath fading experiments. The equal error rate (EER) of the verification experiments with six authorized devices versus six unseen unauthorized devices is as low as 0.63%.

Enhanced forensic speaker verification performance using the ICA-EBM algorithm under noisy and reverberant environments

Forensic speaker verification performance reduces significantly under high levels of noise and reverberation. Multiple channel speech enhancement algorithms, such as independent component analysis by entropy bound minimization (ICA-EBM), can be used to improve noisy forensic speaker verification performance. Although the ICA-EBM was used in previous studies to separate mixed speech signals under clean conditions, the effectiveness of using the ICA-EBM for improving forensic speaker verification performance under noisy and reverberant conditions has not been investigated yet. In this paper, the ICA-EBM algorithm is used to separate the clean speech from noisy speech signals. Features from the enhanced speech are obtained by combining the feature-warped mel frequency cepstral coefficients with similar features extracted from the discrete wavelet transform. The identity vector (i-vector) length normalized Gaussian probabilistic linear discriminant analysis is used as a classifier. The Australian Forensic Voice Comparison and QUT-NOISE corpora were used to evaluate forensic speaker verification performance under noisy and reverberant conditions. Simulation results demonstrate that forensic speaker verification performance based on ICA-EBM improves compared with that of the traditional independent component analysis under different types of noise and reverberation environments. For surveillance recordings corrupted with different types of noise (CAR, STREET and HOME) at − 10 dB signal to noise ratio, the average equal error rate of the proposed method based on ICA-EBM is better than that of the traditional ICA by 12.68% when the interview recordings are kept clean, and 7.25% when the interview recordings have simulated room reverberations.

Mixture linear prediction Gammatone Cepstral features for robust speaker verification under transmission channel noise

In this paper, we present a Mixture Linear Prediction based approach for robust Gammatone Cepstral Coefficients extraction (MLPGCCs). The proposed method provides performance improvement of Automatic Speaker Verification (ASV) using i-vector and Gaussian Probabilistic Linear Discriminant Analysis GPLDA modeling under transmission channel noise. The performance of the extracted MLPGCCs was evaluated using the NIST 2008 database where a single channel microphone recorded conversational speech. The system is analyzed in the presence of different channel transmission noises such as Additive White Gaussian (AWGN) and Rayleigh fading at various Signals to Noise Ratio (SNR) levels. The evaluation results show that the MLPGCCs features are a promising way for the ASV task. Indeed, the speaker verification performance using the MLPGCCs proposed features is significantly improved compared to the conventional Gammatone Frequency Cepstral Coefficients (GFCCs) and Mel Frequency Cepstral Coefficients (MFCCs) features. For speech signals corrupted with AWGN noise at SNRs ranging from (-5 dB to 15 dB), we obtain a significant reduction of the Equal Error Rate (EER) ranging from 9.41% to 6.65% and 3.72% to 1.50%, compared with conventional MFCCs and GFCCs features respectively. In addition, when the test speech signals are corrupted with Rayleigh fading channel we achieve an EER reduction ranging from 23.63% to 7.8% and from 10.88% to 6.8% compared with conventional MFCCs and GFCCs, respectively. We also found that the combination of GFCCs and MLPGCCs gives the highest performance of speaker verification system. The best performance combination achieved is around EER from 0.43% to 0.59% and 1.92% to 3.88%.

Use of Neumann series decomposition to fit the Weighted Euclidean distance and Inner product scoring models in automatic speaker recognition

It is commonly agreed that the main performance indicators in real life biometric applications are: accuracy, calibration quality and computation cost in the case of large databases use. Thus, the Gaussian Probabilistic Linear Discriminant Analysis (GPLDA) has been considered, during the last decade, as one of the most promising scoring techniques in Automatic Speaker Recognition. In this paper, the Algerian Modern Colloquial Arabic Speech Corpus has been used to evaluate a new i-vectors scoring approach based on Neumann series decomposition to fit the Weighted Euclidean Distance and Weighted Inner Product scoring models. This realistic evaluation investigates how this new approach can improve the global recognition performance compared to the GPLDA baseline system, considered as a reference point. The obtained results have shown that the proposed method outperforms the GPLDA baseline system in terms of recognition accuracy and calibration quality. More interestingly, the computational cost of the proposed method is significantly reduced compared to the GPLDA baseline counterpart. These results have an important implication in real-life speaker verification scenarios, especially in some tracking applications, where the computation cost is one of the main concerns besides performance.

Speech Replay Spoofing Attack Detection System Based on Fusion of Classification Algorithms

Fast development of modern technologies of digital processing and speech recording leads to the fact that it is necessary to take into account the potential threats from the speech replay attacks. We propose our ensemble fusion replay attack detection system. It uses constant Q cepstral coefficients as speech features and short-time mean normalization for their preprocessing. The set of binary classifiers includes multiple Gaussian mixture models based Bayesian classifier, i-vector based Gaussian Probabilistic Linear Discriminant Analysis and XGBoost tree boosting algorithm. Fusion of scores was made by modified logistic regression algorithm from BOSARIS toolbox. ASV Spoof 2017 corpus is utilized in the experiments as the main database for anti-spoofing systems evaluation. Obtained results demonstrate that the proposed system can provide substantially better performance than the baseline Gaussian mixture model classifier. The pre-processing of cepstral features is crucial for the better performance of the system. High evaluation performance can be obtained using only few algorithms in a set. The attained value of equal error rate EER=12.44% for our fusion classifier is competitive with the best results obtained during last two years.DOI 10.14258/izvasu(2018)1-19

Joint Estimation of PLDA and Nonlinear Transformations of Speaker Vectors

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.

Exploring kernel discriminant analysis for speaker verification with limited test data

Speaker verification (SV) with limited test data condition is desirable for practical application oriented systems. The i-vector based speaker modeling has shown its significance for SV tasks, but its performance degrades as the utterance becomes shorter. The i-vectors apart from being compact and dominant speaker representations, bear channel and session information, which has to be compensated for robust speaker modeling. The conventional techniques for channel/session compensation include linear discriminant analysis (LDA) followed by within class covariance normalization (WCCN) and Gaussian probabilistic linear discriminant analysis (GPLDA) that eliminate the channel/session variation across the i-vectors by assuming these are linearly separable. In this work, a novel method for channel/session compensation is proposed using kernel discriminant analysis (KDA) that projects the i-vectors into a higher dimensional space and performs discriminant analysis to remove the unwanted information for speaker modeling. The SV studies are performed on standard NIST speaker recognition evaluation (SRE) 2003 and 2008 databases that convey the significance of the proposed compensation over the conventional methods, which is greater on using short test utterances. The achieved improvements are hypothesized due to the non-linearities of channel/session information in the i-vector domain.

A study on the effects of using short utterance length development data in the design of GPLDA speaker verification systems

This paper studies the performance degradation of Gaussian probabilistic linear discriminant analysis (GPLDA) speaker verification system, when only short-utterance data is used for speaker verification system development. Subsequently, a number of techniques, including utterance partitioning and source-normalised weighted linear discriminant analysis (SN-WLDA) projections are introduced to improve the speaker verification performance in such conditions. Experimental studies have found that when short utterance data is available for speaker verification development, GPLDA system overall achieves best performance with a lower number of universal background model (UBM) components. As a lower number of UBM components significantly reduces the computational complexity of speaker verification system, that is a useful observation. In limited session data conditions, we propose a simple utterance-partitioning technique, which when applied to the LDA-projected GPLDA system shows over 8% relative improvement on EER values over baseline system on NIST 2008 truncated 10---10 s conditions. We conjecture that this improvement arises from the apparent increase in the number of sessions arising from our partitioning technique and this helps to better model the GPLDA parameters. Further, partitioning SN-WLDA-projected GPLDA shows over 16% and 6% relative improvement on EER values over LDA-projected GPLDA systems respectively on NIST 2008 truncated 10---10 s interview-interview, and NIST 2010 truncated 10---10 s interview-interview and telephone-telephone conditions.

A General Bayesian Model for Speaker Verification

This paper presents a general Bayesian model for speaker verification tasks. It is a generative probability model. Due to its simple analytical property, a computationally efficient expectation-maximization algorithm can be derived to obtain the model parameters. A closedform solution, which allows the scalable size of enrollment set, is given in a full Bayesian way for making speaker verification decisions. Factor analysis technique is employed to model the speaker-specific components, then the redundant information in this model will be dropped. Experimental results are evaluated by both equal error rate and minimum detection cost function. The proposed approach shows promising results on the National institute of standards and technology (NIST) Speaker recognition evaluation (SRE) 2010 extended and 2012 core tasks. Significant improvement is obtained when comparing with Gaussian probabilistic linear discriminant analysis, especially under phone-call conditions and mismatched train-test channel conditions. Contrast experimental results with other popular generative probability models are also presented in this paper.

Voice biometrics using linear Gaussian model

This study introduces a linear Gaussian model-based framework for voice biometrics. The model works with discrete-time linear dynamical systems. The study motivation is to use the linear Gaussian modelling method in voice biometrics, and show that the accuracy offered by the linear Gaussian modelling method is comparable with other state-of-the-art methods such as Probabilistic Linear Discriminant Analysis and two-covariance model. An expectation-maximisation algorithm is derived to train the model and a Bayesian solution is used to calculate the log-likelihood ratio score of all trials of speakers. This approach performed well on the core-extended conditions of the NIST 2010 Speaker Recognition Evaluation, and is competitive compared with the Gaussian probabilistic linear discriminant analysis, in terms of normalised decision cost function.

Comparison between supervised and unsupervised learning of probabilistic linear discriminant analysis mixture models for speaker verification

We present a comparison of speaker verification systems based on unsupervised and supervised mixtures of probabilistic linear discriminant analysis (PLDA) models. This paper explores current applicability of unsupervised mixtures of PLDA models with Gaussian priors in a total variability space for speaker verification. Moreover, we analyze the experimental conditions under which this application is advantageous, taking into account the existing limitations of training database sizes, provided by the National Institute of Standards and Technology (NIST). We also present a full derivation of the Maximum Likelihood learning procedure for PLDA mixture. Experimental results for a cross-channel NIST Speaker Recognition Evaluation (SRE) 2010 verification task show that unsupervised PLDA mixture is more effective than other state-of-the-art methods. We show that for this task a combination of a homogeneous i-vector extractor and a mixture of two Gaussian PLDA models is more effective than a cross-channel i-vector extractor with a single Gaussian PLDA.

Acoustic Factor Analysis for Robust Speaker Verification

Factor analysis based channel mismatch compensation methods for speaker recognition are based on the assumption that speaker/utterance dependent Gaussian Mixture Model (GMM) mean super-vectors can be constrained to reside in a lower dimensional subspace. This approach does not consider the fact that conventional acoustic feature vectors also reside in a lower dimensional manifold of the feature space, when feature covariance matrices contain close to zero eigenvalues. In this study, based on observations of the covariance structure of acoustic features, we propose a factor analysis modeling scheme in the acoustic feature space instead of the super-vector space and derive a mixture dependent feature transformation. We demonstrate how this single linear transformation performs feature dimensionality reduction, de-correlation, normalization and enhancement, at once. The proposed transformation is shown to be closely related to signal subspace based speech enhancement schemes. In contrast to traditional front-end mixture dependent feature transformations, where feature alignment is performed using the highest scoring mixture, the proposed transformation is integrated within the speaker recognition system using a probabilistic feature alignment technique, which nullifies the need for regenerating the features/retraining the Universal Background Model (UBM). Incorporating the proposed method with a state-of-the-art i-vector and Gaussian Probabilistic Linear Discriminant Analysis (PLDA) framework, we perform evaluations on National Institute of Science and Technology (NIST) Speaker Recognition Evaluation (SRE) 2010 core telephone and microphone tasks. The experimental results demonstrate the superiority of the proposed scheme compared to both full-covariance and diagonal covariance UBM based systems. Simple equal-weight fusion of baseline and proposed systems also yield significant performance gains.