Supervector-based approaches in a discriminative framework for speaker verification in noisy environments

This paper explores the robustness of supervector-based speaker modeling approaches for speaker verification (SV) in noisy environments. In this paper speaker modeling is carried out in two different frameworks: (i) Gaussian mixture model-support vector machine (GMM-SVM) combined method and (ii) total variability modeling method. In the GMM-SVM combined method, supervectors obtained by concatenating the mean of an adapted speaker GMMs are used to train speaker-specific SVMs during the training/enrollment phase of SV. During the evaluation/testing phase, noisy test utterances transformed into supervectors are subjected to SVM-based pattern matching and classification. In the total variability modeling method, large size supervectors are reduced to a low dimensional channel robust vector (i-vector) prior to SVM training and subsequent evaluation. Special emphasis has been laid on the significance of a utterance partitioning technique for mitigating data-imbalance and utterance duration mismatches. An adaptive boosting algorithm is proposed in the total variability modeling framework for enhancing the accuracy of SVM classifiers. Experiments performed on the NIST-SRE-2003 database with training and test utterances corrupted with additive noises indicate that the aforementioned modeling methods outperform the standard GMM-universal background model (GMM-UBM) framework for SV. It is observed that the use of utterance partitioning and adaptive boosting in the speaker modeling frameworks result in substantial performance improvements under degraded conditions.     

Employing both gender and emotion cues to enhance speaker identification performance in emotional talking environments

Speaker recognition performance in emotional talking environments is not as high as it is in neutral talking environments. This work focuses on proposing, implementing, and evaluating a new approach to enhance the performance in emotional talking environments. The new proposed approach is based on identifying the unknown speaker using both his/her gender and emotion cues. Both Hidden Markov Models (HMMs) and Suprasegmental Hidden Markov Models (SPHMMs) have been used as classifiers in this work. This approach has been tested on our collected emotional speech database which is composed of six emotions. The results of this work show that speaker identification performance based on using both gender and emotion cues is higher than that based on using gender cues only, emotion cues only, and neither gender nor emotion cues by 7.22 %, 4.45 %, and 19.56 %, respectively. This work also shows that the optimum speaker identification performance takes place when the classifiers are completely biased towards suprasegmental models and no impact of acoustic models in the emotional talking environments. The achieved average speaker identification performance based on the new proposed approach falls within 2.35 % of that obtained in subjective evaluation by human judges.     

Stochastic feature compensation methods for speaker verification in noisy environments

This paper explores the significance of stereo-based stochastic feature compensation (SFC) methods for robust speaker verification (SV) in mismatched training and test environments. Gaussian Mixture Model (GMM)-based SFC methods developed in past has been solely restricted for speech recognition tasks. Application of these algorithms in a SV framework for background noise compensation is proposed in this paper. A priori knowledge about the test environment and availability of stereo training data is assumed. During the training phase, Mel frequency cepstral coefficient (MFCC) features extracted from a speaker's noisy and clean speech utterance (stereo data) are used to build front end GMMs. During the evaluation phase, noisy test utterances are transformed on the basis of a minimum mean squared error (MMSE) or maximum likelihood (MLE) estimate, using the target speaker GMMs. Experiments conducted on the NIST-2003-SRE database with clean speech utterances artificially degraded with different types of additive noises reveal that the proposed SV systems strictly outperform baseline SV systems in mismatched conditions across all noisy background environments.     

Two-space variability compensation technique for speaker verification in short length and reverberant environments

The performance of state-of-the-art speaker verification in uncontrolled environment is affected by different variabilities. Short duration variability is very common in these scenarios and causes the speaker verification performance to decrease quickly while the duration of verification utterances decreases. Linear discriminant analysis (LDA) is the most common session variability compensation algorithm, nevertheless it presents some shortcomings when trained with insufficient data. In this paper we introduce two methods for session variability compensation to deal with short-length utterances on i-vector space. The first method proposes to incorporate the short duration variability information in the within-class variance estimation process. The second proposes to compensate the session and short duration variabilities in two different spaces with LDA algorithms (2S-LDA). First, we analyzed the behavior of the within and between class scatters in the first proposed method. Then, both proposed methods are evaluated on telephone session from NIST SRE-08 for different duration of the evaluation utterances: full (average 2.5 min), 20, 15, 10 and 5 s. The 2S-LDA method obtains good results on different short-length utterances conditions in the evaluations, with a EER relative average improvement of 1.58%, compared to the best baseline (WCCN[LDA]). Finally, we applied the 2S-LDA method in speaker verification under reverberant environment, using different reverberant conditions from Reverb challenge 2013, obtaining an improvement of 8.96 and 23% under matched and mismatched reverberant conditions, respectively.     

Speaker identification investigation and analysis in unbiased and biased emotional talking environments

This work aims at investigating and analyzing speaker identification in each unbiased and biased emotional talking environments based on a classifier called Suprasegmental Hidden Markov Models (SPHMMs). The first talking environment is unbiased towards any emotion, while the second talking environment is biased towards different emotions. Each of these talking environments is made up of six distinct emotions. These emotions are neutral, angry, sad, happy, disgust and fear. The investigation and analysis of this work show that speaker identification performance in the biased talking environment is superior to that in the unbiased talking environment. The obtained results in this work are close to those achieved in subjective assessment by human judges.     

Speaker identification in emotional talking environments based on CSPHMM2s

Speaker recognition systems perform almost ideal in neutral talking environments; however, these systems perform poorly in emotional talking environments. This research is devoted to enhancing the low performance of text-independent and emotion-dependent speaker identification in emotional talking environments based on employing Second-Order Circular Suprasegmental Hidden Markov Models (CSPHMM2s) as classifiers. This work has been tested on our speech database which is composed of 50 speakers talking in six different emotional states. These states are neutral, angry, sad, happy, disgust, and fear. Our results show that the average speaker identification performance in these talking environments based on CSPHMM2s is 81.50% with an improvement rate of 5.61%, 3.39%, and 3.06% compared, respectively, to First-Order Left-to-Right Suprasegmental Hidden Markov Models (LTRSPHMM1s), Second-Order Left-to-Right Suprasegmental Hidden Markov Models (LTRSPHMM2s), and First-Order Circular Suprasegmental Hidden Markov Models (CSPHMM1s). Our results based on subjective evaluation by human judges fall within 2.26% of those obtained based on CSPHMM2s.     

A novel scores fusion approach applied on speaker verification under noisy environments

To improve the speaker verification system in adverse conditions, a novel score fusion approach using adaptive method, based on a prior Equal Error Rate (EER), is presented in this paper. Currently, the most commonly used methods are the mean, product, minimum, maximum, or the weighted sum of scores. Our method introduces the MLP network which approximates the estimated scores under noisy conditions, to those of the ideal estimated in clean environments and gives the optimally weighted parameters, to be added in the adaptive weights used for weighting sum of scores. This method is assessed by using the NIST 2000 corpus and different feature extraction methods. Noisy conditions are created using NOISEX-92. In severely degraded conditions, the results show that the speaker verification process using our proposed score fusion approach applied to the GMM-UBM and GMM-SVM based systems, achieves better performances in terms of EER reduction than each system used alone.     

Improving the performance of GPLDA speaker verification using unsupervised inter-dataset variability compensation approaches

In practical applications, speaker verification systems have to be developed and trained using data which is outside the domain of the intended application as the collection of significant amount of in-domain data could be difficult. Experimental studies have found that when a GPLDA system is trained using out-domain data, it significantly affects the speaker verification performance due to the mismatch between development data and evaluation data. This paper proposes several unsupervised inter-dataset variability compensation approaches for the purpose of improving the performance of GPLDA systems trained using out-domain data. We show that when GPLDA is trained using out-domain data, we can improve the performance by as much as 39% by using by score normalisation using small amounts of in-domain data. Also in situations where rich out-domain data and only limited in-domain data are available, a pooled-linear-weighted technique to estimate the GPLDA parameters shows 35% relative improvements in equal error rate (EER) on int–int conditions. We also propose a novel inter-dataset covariance normalization (IDCN) approach to overcome in- and out-domain data mismatch problem. Our unsupervised IDCN-compensated GPLDA system shows 14 and 25% improvement respectively in EER over out-domain GPLDA speaker verification on tel–tel and int–int training–testing conditions. We provide intuitive explanations as to why these inter-dataset variability compensation approaches provide improvements to speaker verification accuracy.     

Sample-specific late classifier fusion for speaker verification

Multimedia Tools and Applications - Due to the mismatch between training and test conditions, speaker verification in real environments, continues to be a challenging problem. An effective way of...     

Local spectral variability features for speaker verification

Speaker verification techniques neglect the short-time variation in the feature space even though it contains speaker related attributes. We propose a simple method to capture and characterize this spectral variation through the eigenstructure of the sample covariance matrix. This covariance is computed using sliding window over spectral features. The newly formulated feature vectors representing local spectral variations are used with classical and state-of-the-art speaker recognition systems. Results on multiple speaker recognition evaluation corpora reveal that eigenvectors weighted with their normalized singular values are useful in representing local covariance information. We have also shown that local variability features can be extracted using mel frequency cepstral coefficients (MFCCs) as well as using three recently developed features: frequency domain linear prediction (FDLP), mean Hilbert envelope coefficients (MHECs) and power-normalized cepstral coefficients (PNCCs). Since information conveyed in the proposed feature is complementary to the standard short-term features, we apply different fusion techniques. We observe considerable relative improvements in speaker verification accuracy in combined mode on text-independent (NIST SRE) and text-dependent (RSR2015) speech corpora. We have obtained up to 12.28% relative improvement in speaker recognition accuracy on text-independent corpora. Conversely in experiments on text-dependent corpora, we have achieved up to 40% relative reduction in EER. To sum up, combining local covariance information with the traditional cepstral features holds promise as an additional speaker cue in both text-independent and text-dependent recognition.     

Constrained temporal structure for text-dependent speaker verification

In the context of mobile devices, speaker recognition engines may suffer from ergonomic constraints and limited amount of computing resources. Even if they prove their efficiency in classical contexts, GMM/UBM systems show their limitations when restricting the quantity of speech data. In contrast, the proposed GMM/UBM extension addresses situations characterised by limited enrolment data and only the computing power typically found on modern mobile devices. A key contribution comes from the harnessing of the temporal structure of speech using client-customised pass-phrases and new Markov model structures. Additional temporal information is then used to enhance discrimination with Viterbi decoding, increasing the gap between client and imposter scores. Experiments on the MyIdea database are presented with a standard GMM/UBM configuration acting as a benchmark. When imposters do not know the client pass-phrase, a relative gain of up to 65% in terms of EER is achieved over the GMM/UBM baseline configuration. The results clearly highlight the potential of this new approach, with a good balance between complexity and recognition accuracy.     

A nonlinear autoregressive model for speaker verification

Gaussian Mixture Models (GMM) have been the most popular approach in speaker recognition and verification for over two decades. The inefficiencies of this model for signals such as speech are well documented and include an inability to model temporal dependencies that result from nonlinearities in the speech signal. The resulting models are often complex and overdetermined, which leads to a lack of generalization. In this paper, we present a nonlinear mixture autoregressive model (MixAR) that attempts to directly model nonlinearities in the trajectories of the speech features. We apply this model to the problem of speaker verification. Experiments with synthetic data demonstrate the viability of the model. Evaluations on standard speech databases, including TIMIT, NTIMIT, and NIST-2001, demonstrate that MixAR, using only half the number of parameters and only static features, can achieve a lower equal error rate when compared to GMMs, particularly in the presence of previously unseen noise. Performance as a function of the duration of both the training and evaluation utterances is also analyzed.     

Frame level sparse representation classification for speaker verification

Multimedia Tools and Applications - In this paper, we analyze the application of the sparse representation of frames of the speech signal for the speaker verification. It is lately shown that...     

Articulatory movement features for short-duration text-dependent speaker verification

During our pronunciation process, the position and movement properties of articulators such as tongue, jaw, lips, etc are mainly captured by the articulatory movement features (AMFs). This paper investigates to use the AMFs for short-duration text-dependent speaker verification. The AMFs can characterize the relative motion trajectory of articulators of individual speakers directly, which is rarely affected by the external environment. Therefore, we expect that, the AMFs are superior to the traditional acoustic features, such as mel-frequency cepstral coefficients (MFCC), to characterize the speaker identity differences between speakers. The speaker similarity scores measured by the dynamic time warping (DTW) algorithm are used to make the speaker verification decisions. Experimental results show that the AMFs can bring significant performance gains over the traditional MFCC features for short-duration text-dependent speaker verification task.     

Quality measures for speaker verification with short utterances

The performances of the automatic speaker verification (ASV) systems degrade due to the reduction in the amount of speech used for enrollment and verification. Combining multiple systems based on different features and classifiers considerably reduces speaker verification error rate with short utterances. This work attempts to incorporate supplementary information during the system combination process. We use quality of the estimated model parameters as supplementary information. We introduce a class of novel quality measures formulated using the zero-order sufficient statistics used during the i-vector extraction process. We have used the proposed quality measures as side information for combining ASV systems based on Gaussian mixture model–universal background model (GMM–UBM) and i-vector. The proposed methods demonstrate considerable improvement in speaker recognition performance on NIST SRE corpora, especially in short duration conditions. We have also observed improvement over existing systems based on different duration-based quality measures.     

Manifold learning based speaker dependent dimension reduction for robust text independent speaker verification

Speaker verification has been studied widely from different points of view, including accuracy, robustness and being real-time. Recent studies have turned toward better feature stability and robustness. In this paper we study the effect of nonlinear manifold based dimensionality reduction for feature robustness. Manifold learning is a popular recent approach for nonlinear dimensionality reduction. Algorithms for this task are based on the idea that each data point may be described as a function of only a few parameters. Manifold learning algorithms attempt to uncover these parameters in order to find a low-dimensional representation of the data. From the manifold based dimension reduction approaches, we applied the widely used Isometric mapping (Isomap) algorithm. Since in the problem of speaker verification, the input utterance is compared with the model of the claiming client, a speaker dependent feature transformation would be beneficial for deciding on the identity of the speaker. Therefore, our first contribution is to use Isomap dimension reduction approach in the speaker dependent context and compare its performance with two other widely used approaches, namely principle component analysis and factor analysis. The other contribution of our work is to perform the nonlinear transformation in a speaker-dependent framework. We evaluated this approach in a GMM based speaker verification framework using Tfarsdat Telephone speech dataset for different noises and SNRs and the evaluations have shown reliability and robustness even in low SNRs. The results also show better performance for the proposed Isomap approach compared to the other approaches.     

Processing degraded speech for text dependent speaker verification

This work explores the use of speech enhancement for enhancing degraded speech which may be useful for text dependent speaker verification system. The degradation may be due to noise or background speech. The text dependent speaker verification is based on the dynamic time warping (DTW) method. Hence there is a necessity of the end point detection. The end point detection can be performed easily if the speech is clean. However the presence of degradation tends to give errors in the estimation of the end points and this error propagates into the overall accuracy of the speaker verification system. Temporal and spectral enhancement is performed on the degraded speech so that ideally the nature of the enhanced speech will be similar to the clean speech. Results show that the temporal and spectral processing methods do contribute to the task by eliminating the degradation and improved accuracy is obtained for the text dependent speaker verification system using DTW.     

说话人确认中分数归一化的一种新方法

在说话人确认中，由于目标说话人和冒认者的得分分布是双峰分布，并且不同目标说话人模型得分分布不一致，使对所有说话人确定一个统一的阈值变得困难，导致系统性能下降。分数归一化通过调整冒认者的得分分布来调整阈值。简要介绍了目前最常用的两种归一化方法：零归一化（Z-Norm）和测试归一化(T-Norm)。重点引入了一种新的根据KL距离的D-Norm 归一化方法。然后结合Z-Norm 和D-Norm的优点，又提出一种新的方法ZD-Norm。对这四种归一化方法的性能进行了比较。实验表明，ZD-Norm相对Z-Norm和D-Norm，能够更有效地提高说话人确认系统的性能。