Speaker recognition using dynamic synapse neural networks

Synaptic dynamics: Linear model and adaptation algorithm

Identification of acoustic signals in noisy environments remains one of the most difficult of signal processing problems, and is a major obstacle to the high degree of accuracy and speed needed to identify suspicious sounds in high-security, high-safety environments. We have previously developed an acoustic recognition capability using a novel, biologically based Dynamic Synapse Neural Network (DSNN) technology. The DSNN-based technology has been demonstrated to classify target sounds with a high degree of accuracy, even in high noise conditions. In this paper we focus on extending the acoustic recognition capability of DSNNs to the problem of gunshot recognition. In order to recognize and localize the event, an array of four microphones is used for sound input. For localization purpose, time-delay estimation algorithm (TDE) is employed for triangulation. We have developed stand-alone, portable, and cost-efficient hardware with which both recognition and localization can be performed. In field-testing, the system classifies and localizes over 90% of the trained-for sounds. Performance for firecracker, starter pistol, 9-mm, and 44-caliber, explosion/firing sounds was also tested.

This paper deals with the adaptive nonlinear identification and trajectory tracking via dynamic multilayer neural network (NN) with different timescales. Two NN identifiers are proposed for nonlinear systems identification via dynamic NNs with different timescales including both fast and slow phenomenon. The first NN identifier uses the output signals from the actual system for the system identification. In the second NN identifier, all the output signals from nonlinear system are replaced with the state variables of the NNs. The online identification algorithms for both NN identifier parameters are proposed using Lyapunov function and singularly perturbed techniques. With the identified NN models, two indirect adaptive NN controllers for the nonlinear systems containing slow and fast dynamic processes are developed. For both developed adaptive NN controllers, the trajectory errors are analyzed and the stability of the systems is proved. Simulation results show that the controller based on the second identifier has better performance than that of the first identifier.

Recurrent neural networks unlike feed-forward networks are able to process inputs with time context. The key role in this process is played by the dynamics of the network, which transforms input data to the recurrent layer states. Several authors have described and analyzed dynamics of small sized recurrent neural networks with two or three hidden units. In our work we introduce techniques that allow to visualize and analyze the dynamics of large recurrent neural networks with dozens units, reveal both stable and unstable points (attractors and saddle points), which are important to understand the principles of successful task processing. As a practical example of this approach, dynamics of the simple recurrent network trained by two different training algorithms on context-free language anbnwas studied.

При тестировании на специализированных базах данных наиболее совершенных систем идентификации диктора их минимальная эффективность, оцениваемая величиной вероятности ошибки в точке пересечения кривых ошибок, составляет лишь несколько процентов. Однако известно множество факторов, влияющих на вариативность характеристик голоса диктора, каждый из которых имеет свое, отличное от других, влияние на результаты идентификации диктора по характеристикам голоса. Сложность создания и тестирования систем идентификации диктора заключается в необходимости количественной формализации ряда конкретных факторов, влияющих на характеристики его голоса. Рассмотрен предложенный метод учета множества факторов, влияющих на параметры характеристик голоса диктора, обеспечивающий принципиальную возможность косвенного учета их практически неограниченного количества. Согласно этому методу из речевых сигналов выделяются «атомарные» структуры, зависящие от совокупности основных факторов, влияющих на процесс идентификации диктора. По такому методу все существенные факторы, влияющие на характеристики голоса, будут косвенно учитываться на уровне этих структур. Экспертные решения принимаются по комбинаторной совокупности огромного числа этих «атомарных» структур. Под «атомарными» структурами речи понимаются спектры любых фрагментов гласных звуков, выделяемых во временном окне продолжительностью 20 мс. "Атомарные" структуры выделяются в автоматическом режиме. Предлагаемый метод обеспечивает рациональный учет богатофакторности влияния различных параметров, поскольку на спектры этих структур влияют все основные факторы, характеризующие индивидуальность голоса конкретного диктора. Решение об идентичности голосов дикторов, записанных на разных фонограммах, осуществляется на основе комбинаторики «атомарных» спектров гласных звуков в обоих фонограммах. Метод показал высокую эффективность при экспертизе фонограмм малой продолжительности.

One of the difficulties encountered in the application of reinforcement learning methods to real-world problem is the generalization of large-scale or continuous state space. In order to solve the curse of dimensionality problem caused by discretizing continuous state space, a kind of Q-learning method for continuous state space based on a dynamic Elman neural network was proposed in this paper. The inputs and the output of Elman network are the system state-action pair and the corresponding Q-value. That is, Elman network is used to estimate the Q-value of state-action pair on-line. Eligibility trace for connecting weights is introduced by borrowing ideas from the eligibility trace mechanism of state in temporal difference algorithm to improve the learning speed of neural network. Computer simulations on mountain car control illustrate the performance and applicability of the proposed Q-learning scheme.

Several mathematical approaches to studying analytically the dynamics of neural networks rely on mean-field approximations, which are rigorously applicable only to networks of infinite size. However, all existing real biological networks have finite size, and many of them, such as microscopic circuits in invertebrates, are composed only of a few tens of neurons. Thus, it is important to be able to extend to small-size networks our ability to study analytically neural dynamics. Analytical solutions of the dynamics of small-size neural networks have remained elusive for many decades, because the powerful methods of statistical analysis, such as the central limit theorem and the law of large numbers, do not apply to small networks. In this article, we critically review recent progress on the study of the dynamics of small networks composed of binary neurons. In particular, we review the mathematical techniques we developed for studying the bifurcations of the network dynamics, the dualism between neural activity and membrane potentials, cross-neuron correlations, and pattern storage in stochastic networks. Then, we compare our results with existing mathematical techniques for studying networks composed of a finite number of neurons. Finally, we highlight key challenges that remain open, future directions for further progress, and possible implications of our results for neuroscience.

In this paper a Dynamic Neural Network has been designed to effectively estimate the states of a nonlinear single input plant. The Unscented Kalman Filter algorithm has been used to train the weights of the Dynamic Neural Network. No a priori bounds of the weights and errors are required. The non linear plant and the Dynamic Neural Network model have been simulated by the same input to illustrate the validity of the theoretical results.

Traditional speaker authentication focuses on speaker verification (SV) and speaker identification, which is accomplished by matching the speaker's voice with his or her registered speech patterns. In this paper, we propose a new technique, verbal information verification (VIV), in which spoken utterances of a claimed speaker are verified against the key (usually confidential) information in the speaker's registered profile automatically; to decide whether the claimed identity should be accepted or rejected. Using the proposed sequential procedure involving three question-response turns, we achieved an error-free result in a telephone speaker authentication experiment with 100 speakers. We further propose a speaker authentication system by combining VIV with SV. In the system, a user is verified by VIV in the first four to five accesses, usually from different acoustic environments. During these uses, one of the key questions pertains to a pass-phrase for SV. The VIV system collects and verifies the pass-phrase utterance for use as training data for speaker model construction. After a speaker-dependent model is constructed, the system then migrates to SV. This approach avoids the inconvenience of a formal enrollment procedure, ensures the quality of the training data for SV, and mitigates the mismatch caused by different acoustic environments between training and testing. Experiments showed that the proposed system improved the SV performance by over 40% in equal-error rate compared to a conventional SV system.

With the ubiquity of complex networks in domains as diverse as social network analysis, recommender systems and epidemiology, improvements in link predictions have a potential for far-reaching impact. In machine learning, a well-known approach to improve classification performance is to combine several methods in an ensemble. It is also well established that, diversity or complementarity of the classifiers to be combined, is a prerequisite for a potential performance improvement of the formed ensemble. In this work, we combine link prediction heuristics (e.g. Common Neighbour, Adamic-Adar etc.), with static graph neural networks, discrete dynamic graph neural networks, and continuous dynamic graph neural networks. We analyze the importance of each of the methods and how complementary they are with respect to each other. We also use greedy searches to select better subsets of methods to combine and train the ensemble in a roll-forward manner to make it adapt to changes in the dynamic network over time. In the conducted experiments on a number of challenging link prediction tasks, the ensemble from the greedy search and the roll-forward ensemble shows on average a 55% and 61% improvement respectively, over the best single method.

Data efficient voice cloning aims at synthesizing target speaker's voice with only a few enrollment samples at hand. To this end, speaker adaptation and speaker encoding are two typical methods based on base model trained from multiple speakers. The former uses a small set of target speaker data to transfer the multi-speaker model to target speaker's voice through direct model update, while in the latter, only a few seconds of target speaker's audio directly goes through an extra speaker encoding model along with the multi-speaker model to synthesize target speaker's voice without model update. Nevertheless, the two methods need clean target speaker data. However, the samples provided by user may inevitably contain acoustic noise in real applications. It's still challenging to generating target voice with noisy data. In this paper, we study the data efficient voice cloning problem from noisy samples under the sequence-to-sequence based TTS paradigm. Specifically, we introduce domain adversarial training (DAT) to speaker adaptation and speaker encoding, which aims to disentangle noise from speech-noise mixture. Experiments show that for both speaker adaptation and encoding, the proposed approaches can consistently synthesize clean speech from noisy speaker samples, apparently outperforming the method adopting state-of-the-art speech enhancement module.

Biological modeling the undulatory locomotion of C. elegans using dynamic neural network approach

Dynamic neural networks can greatly reduce computation redundancy without compromising accuracy by adapting their structures based on the input. In this paper, we explore the robustness of dynamic neural networks against energy-oriented attacks targeted at reducing their efficiency. Specifically, we attack dynamic models with our novel algorithm GradMDM. GradMDM is a technique that adjusts the direction and the magnitude of the gradients to effectively find a small perturbation for each input, that will activate more computational units of dynamic models during inference. We evaluate GradMDM on multiple datasets and dynamic models, where it outperforms previous energy-oriented attack techniques, significantly increasing computation complexity while reducing the perceptibility of the perturbations https://github.com/lingengfoo/GradMDM.

The use of computers has increased to a very large extent due to its immense use and growing technologies that gives rise to transfer of digital media over longer distances. As the audio data may travel over large networks it needs to be purified or cleaned during recording or while the processing the data so that it can be identified. Cleaning is related to removing the noise that may exist in the signal. There are 2 ways in which the audio can be cleaned. First, recording can be done in clean environment that includes no noise at all, even the fans may create disturbances, recording done in a closed room. But feature extraction in this case is more suitable using MFCC. Second, if recording is already done and needs to processed and if it contains lot of noise than some noise reduction technique such as wavelet or filter can used to eliminate the noise. In this paper, we will implement Gamma tone Frequency Cepstral Coefficient (GFCC) to eliminate the noise components. This is due to the fact that MFCC does not give required accuracy in the presence of noise. It works well in clean environment. On the other hand, GFCC gives the required performance and accuracy in clean as well as in the presence of noise. This paper compares auditory feature based methods namely GFCC and MFCC and concludes which method is more suitable in which environment. This resulting GFCC features are used by k- means clustering to group the similar voices in to clusters and thus identify the singer. Dynamic time warping is used to align or match two sequences that may vary in time or speed. It is a pattern matching technique that will authenticate the person a claimed identity.

The article is devoted to the problem of optimization of search request ranking algorithms in the digital information retrieval system. The algorithm of functioning of the neural network ranking unit based on Hopfield neural network is built. The ability to generate a ranked list of pages found as a result of the request in the digital information retrieval system can be provided by solving two problems of integer optimization: the problem of assignment of combinatorial sets of criteria for assessing the relevance of web page search and the problem of sorting of numbers—relevance values. The architecture of the neural network model based on the dynamic Hopfield neural network with binary output function designed for combinatorial optimization of the final list of documents found in the digital information retrieval system was synthesized. Promising variants of neural network models with binary output function of neurons for synthesis of the optimal evaluation plan with a combinatorial set of criteria by solving the problem of assignment were built. It has been proven that the built models differ in the rules for determining the coefficients of synaptic connections and external shifts; each of the created rules can be used independently or in different combinations with one another. In the course of analytical research, it was found that the optimization formulation of the problem of sorting of relevance values of search pages is identical to the problem of assignment of combinatorial groups of evaluation criteria provided that the elements of the performance matrix of the latter are defined as linear combinations of relevance values.