Generalizing I-Vector Estimation for Rapid Speaker Recognition

The design and assessment of structural dynamics behaviour often rely on the ‘modal properties’ (primarily modal frequencies, damping ratios and mode shapes) of a structure. Modal properties are frequently identified in vibration related applications as their information can benefit design theory validation and in-situ property updating and monitoring. Given the prospect of flexibility and high implementation economy, operational modal analysis (OMA) aims to identify the modal properties using only ambient vibration response whose amplitude can be small. When instrument noise becomes a non-negligible part of measured data, the identified result can have significant uncertainty, calling for proper uncertainty quantification. Amidst the complex composition of ambient vibration data of civil structures, conventional noise assumption may not always apply. Motivated by issues about uncertainty modelling and computation, this thesis contributes to the theory and computational algorithm of OMA from a Bayesian perspective in the frequency domain. Focusing on the resonance band of vibration modes, a Bayesian probabilistic approach provides a fundamental mathematical framework for OMA. The ‘posterior’ (i.e., given data) most probable value (MPV) of modal properties maximising the likelihood of data informs about the best estimate. The uncertainty associated with the estimate is quantified by the posterior covariance matrix, which can be calculated as the inverse of the Hessian of negative log-likelihood function (NLLF) at the MPV. Existing formulations assume independent and identically distributed (i.i.d.) prediction errors among different channels, which account for both instrument noise and modelling error. While the assumption significantly simplifies theory and computation, and may not matter when the signal-to-noise (s/n) ratio is high, it is prone to be violated in OMA. For example, large disparity among noise intensities may exist, and significant correlation can occur because of the presence of unmodelled dynamics, which may affect identification results. To understand and manage the consequence of these potential violations, a comprehensive investigation on the effect of prediction error models with different statistical assumptions is presented in this work. Beyond the i.i.d. assumption, two generalised prediction error models have been considered to account for the disparity and correlation that may arise in applications. Such investigation was previously hindered by computational issues because the general prediction error models feature more parameters carrying somewhat different mathematical roles from the existing ones, making the computation of posterior MPV and covariance matrix more challenging in the first place. As a contribution to computation and enabling scientific investigation, new efficient algorithms applicable to general prediction error models are developed in this work for their computation. As a contribution to theory, a mathematical analogy for general noise is established with i.i.d. model for OMA problems. Not only does this provide a simple approximate way for handling general noise with existing tools based on i.i.d. assumption, it also allows the understanding and experience accumulated with i.i.d. model to be generalised. As a spin-off of this opportunity, asymptotic formulae for identification uncertainty are developed for general prediction error models in this work, which yield a generalised definition for s/n ratio that provides a fundamental quantification of data quality in OMA. The theory and algorithms are developed and tested against synthetic, laboratory and field data. As a further contribution to computation, motivated by existing overheads and hurdles in computer code development, this work proposes an alternative semi-analytical means for calculating the Hessian of NLLF in OMA. The proposed method explores a mathematical identity in the theory of expectation-maximisation (EM) algorithm to allow the Hessian matrix to be calculated via Monte Carlo simulation, which offers an alternative means for counter-checking/benchmarking during computer code development. Theoretical implications of the identity are also explored, and numerical examples are presented to illustrate implementation aspects.

Even though speaker recognition has gained significant progress in recent years, its performance is known to be deteriorated severely with the existence of strong background noises. Inspired by a recently proposed clean-frame selection approach, this work investigates a relatively elegant weighting method when computing the Baum-Welch statistics of Gaussian mixture models (GMMs) in i-vector extraction. By introducing weighting parameters to the frames of enrollment/testing utterances, the optimization problem is redefined and solved. New updating rules are derived by incorporating weights to the computation of posterior probabilities, mean vectors, and covariance matrices of the GMM. The experiments conducted on the Speakers in the Wild (SITW) database show that the proposed algorithm has significantly improved the performance of i-vector-based speaker recognition systems in noisy environments. Compared with the GMM i-vector baseline, the equal error rate is reduced from 5.75 to 4.72 and the minimum value of cost function (C det min ) is reduced from 0.4825 to 0.4505. Slight but significant superiority is also observed over the method with an additional feature enhancement frontend by using deep neural networks.

We propose a strategy for discriminative training of the ivector extractor in speaker recognition. The original i-vector extractor training was based on the maximum-likelihood generative modeling, where the EM algorithm was used. In our approach, the i-vector extractor parameters are numerically optimized to minimize the discriminative cross-entropy error function. Two versions of the i-vector extraction are studied—the original approach as defined for Joint Factor Analysis, and the simplified version, where orthogonalization of the i-vector extractor matrix is performed. Index Terms: speaker verification, i-vectors, PLDA, discriminative training

Speaker identification refers to the process of recognizing human voice using artificial intelligence techniques. Speaker identification technologies are widely applied in voice authentication, security and surveillance, electronic voice eavesdropping, and identity verification. In the speaker identification process, extracting discriminative and salient features from speaker utterances is an important task to accurately identify speakers. Various features for speaker identification have been recently proposed by researchers. Most studies on speaker identification have utilized short-time features, such as perceptual linear predictive (PLP) coefficients and Mel frequency cepstral coefficients (MFCC), due to their capability to capture the repetitive nature and efficiency of signals. Various studies have shown the effectiveness of MFCC features in correctly identifying speakers. However, the performances of these features degrade on complex speech datasets, and therefore, these features fail to accurately identify speaker characteristics. To address this problem, this study proposes a novel fusion of MFCC and time-based features (MFCCT), which combines the effectiveness of MFCC and time-domain features to improve the accuracy of text-independent speaker identification (SI) systems. The extracted MFCCT features were fed as input to a deep neural network (DNN) to construct the speaker identification model. Results showed that the proposed MFCCT features coupled with DNN outperformed existing baseline MFCC and time-domain features on the LibriSpeech dataset. In addition, DNN obtained better classification results compared with five machine learning algorithms that were recently utilized in speaker recognition. Moreover, this study evaluated the effectiveness of one-level and two-level classification methods for speaker identification. The experimental results showed that two-level classification presented better results than one-level classification. The proposed features and classification model for identifying a speaker can be widely applied to different types of speaker datasets.

Gaussian Mixture models is one of the most popular statistical methods in speaker recognition. The purpose of this research is to perform experiments of speaker recognition using various feature vectors: four formants, four formants with fundamental frequency and mel cepstrum coefficients. Gaussian mixture models using mel cepstrum coefficients is baseline in speaker recognition and gives one of the best results in text independent speaker recognition. After implementing experiments of speaker recognition and comparing experimental results we can affirm that mel scale cepstral coefficients and four formants with fundamental frequency gives quite the same recognition accuracy, but creating of Gaussian mixture speaker models and recognition process continues a few times longer using mel scale cepstral coefficients, because count of calculations is few times greater in that case. Using only four formants gives the worst results of recognition accuracy. Ill. 7, bibl. 12 (in English; summaries in English, Russian and Lithuanian).

Most of the state-of-the-art speaker recognition systems use i-vectors, a compact representation of spoken utterances. Since the “standard” i-vector extraction procedure requires large memory structures, we recently presented the Factorized Sub-space Estimation (FSE) approach, an efficient technique that dramatically reduces the memory needs for i-vector extraction, and is also fast and accurate compared to other proposed approaches. FSE is based on the approximation of the matrix T, representing the speaker variability sub-space, by means of the product of appropriately designed matrices. In this work, we introduce and evaluate a further approximation of the matrices that most contribute to the memory costs in the FSE approach, showing that it is possible to obtain comparable system accuracy using less than a half of FSE memory, which corresponds to more than 60 times memory reduction with respect to the standard method of i-vector extraction.

Speaker recognition using high-level features has been a successful area of exploration. Features obtained from many different levels?phones, words, prosodic events, etc.?are used to characterize the speaker. A good modeling technique for these features is the support vector machine (SVM). SVMs model the n-gram frequencies from speaker utterances in a high-dimensional SVM feature space and have shown excellent performance over a wide variety of high-level features. A complimentary method of recent exploration in SVM speaker recognition is the use of nuisance attribute projection (NAP). NAP removes directions from SVM feature space that are superfluous to the task of speaker recognition?channel information, session variability, etc. In this paper, we consider the application of NAP to high-level speaker recognition. We describe the difficulties in applying this method and propose solutions. We also conduct experiments showing that NAP can reduce variability in SVM feature space leading to improved performance.

In speaker identification, most of the computational processing time is required to calculate the likelihood of the test utterance of the unknown speaker with respect to the speaker models in the database. When number of speakers in the database is in the order of 10,000 or more, then computational complexity becomes very high. In this paper, we propose a Maximum Likelihood Linear Regression (MLLR) based fast method to calculate the likelihood from the speaker model using the MLLR matrix. The proposed technique will help to quickly find the best N speakers during identification. After that final speaker identification task can be done within the N selected speakers using any conventional method of speaker identification. The comparative study of the proposed method is done in terms of processing time with the state-of-the-art GMM-UBM based system on NIST 2004 SRE. The proposed technique performs faster than GMM-UBM based system with some degradation in system accuracy.

Fast Bayesian ambient modal identification in the frequency domain, Part II: Posterior uncertainty

To solve the complex scenarios, such as multimodal noise, bad measurement data and sudden load changes, in the power system, the enhanced generalised cross correlation entropy unscented Kalman filter (EnGCCE‐UKF) method is proposed in this paper. This method replaces the mean square error (MSE) criterion of the traditional UKF with the generalised cross correlation entropy (GCCE) criterion and obtains the optimal solution by optimising the state estimation cost function, significantly improving the robustness and estimation accuracy in the non‐Gaussian noise environment. Considering the interference of bad measurement data on the information matrix, the strong filtering tracking (SFT) theory is further embedded in the GCCE‐UKF framework. By dynamically adjusting the information matrix to suppress the influence of abnormal factors, the anti‐interference ability of the algorithm is enhanced. This method integrates the state and measurement error into the cost function of the EnGCCE by using the statistical linearisation technique and recursively updates the posterior estimation and covariance matrix with the help of the fixed‐point iterative algorithm. Verified through multiscenario experiments and comparative analysis, the proposed method has shown good effectiveness in power systems of three scales.

The cognition radar utilizes the dynamic waveform configuration to adapt to the environment to improve tracking performance. A novel tracking algorithm based on waveform selection is proposed to address the maneuvering aircraft tracking problem in clutter. Based on the modified current statistical (MCS) model, the modified probabilistic data association filter (MPDAF) is integrated with the square-root cubature Kalman filter (SCKF) to deal with the nonlinear measurement and the clutter. Additionally, the waveform library is established by applying the fractional Fourier transform (FRFT) to a base waveform, and a direct method is proposed to select the optimal waveform to minimize the posterior state error covariance matrix. The simulation results showed that, compared with the filters with the fixed waveform as well as the state-of-the-art algorithms, the proposed algorithm achieved higher estimation precision and lower track loss percentage while maintaining a low computational burden.

Image reconstruction algorithm of electromagnetic tomography based on fractional Kalman filter

Maximum correntropy delay Kalman filter for SINS/USBL integrated navigation

In view of abnormal situations, such as multimodal noise, bad measurement data, and sudden load changes in a ship integrated electric propulsion system, this article proposes an unscented Kalman filter with enhanced mixture minimum error entropy (EnMMEE-UKF) to estimate the dynamic state of a ship integrated electric propulsion system. The new algorithm mixture includes two Gaussian functions, defines them as kernel functions, and uses quantization operators to lighten the weight of the mixture minimum error entropy (MMEE). This not only reduces the calculation burden of the MMEE but also effectively improves the system’s ability to handle strongly nonlinear models. In addition, in the lightweight MMEE (LMMEE)-unscented Kalman filter (UKF) algorithm, a fading factor is introduced, based on the strong tracking filter (STF) theory, to modify the error covariance matrix. The enhanced MMEE replaces the mean square error loss of the traditional UKF algorithm. This enhances the robustness of the algorithm estimation, in which a statistical linearization method is adopted to merge the state and measurement errors into the cost function of the enhanced MMEE. Finally, a fixed-point iteration method is applied to recursively update the posterior estimate and covariance matrix. To verify the effectiveness of the proposed method, numerical simulation experiments of various scenarios are conducted on the simulation system and the prototype system. The results demonstrate that the proposed algorithm not only has high estimation accuracy under various abnormal conditions but also effectively reduces the computational complexity of the algorithm, making its application in engineering possible.

Speaker Recognition is the ability of the system to recognize the speaker from the set of speaker samples available in the system. It is of 2 types, one uses a keyword, called text-dependent systems, and another one can recognize the voice in any language/text, also called as text-independent speaker recognition. In this paper, a text-independent, language-independent speaker recognition system is implemented using dense & convolutional neural networks. Speaker recognition has found several applications in upcoming electronic products like personal/home assistants, telephone banking and biometric identification. In this paper, we explore a system that uses MFCC along with DNN and CNN as the model for building a speaker recognition system.