Voice cloning technology has made significant strides in recent years, with applications ranging from personalized virtual assistants to sophisticated entertainment systems. This study compares nine voice cloning models, focusing on both zero-shot and fine-tuned approaches. Zero-shot voice cloning models have gained attention for their ability to generate high-quality synthetic voices without requiring extensive training data for each new voice and for their capability to perform real-time inference online. In contrast, non-zero- shot models typically require additional data but can offer improved fidelity in voice reproduction. The study comprises two key experiments. The first experiment evaluates the performance of zero-shot voice cloning models, analyzing their ability to reproduce target voices without prior exposure accurately. The second experiment involves fine-tuning the models on target speakers to assess improvements in voice quality and adaptability. The models are evaluated based on key metrics assessing voice quality, speaker identity preservation, and subjective and objective performance measures. The findings indicate that while zero-shot models offer greater flexibility and ease of deployment, fine-tuned models can deliver superior performance.

The article deals with the urgent problem of software security at the early stage of its development. Special attention is paid to static code analysis, which is a key tool for detecting vulnerabilities at early stages of the software development life cycle. The article emphasizes the importance of integrating static analysis tools into the development process in order to detect and eliminate vulnerabilities early. The methods of static analyzers’ error search are considered, as well as the main components of the Svace static analyzer developed at the Institute of System Programming of the Russian Academy of Sciences. Classification of analyses used in the Svace static analyzer is presented. Static analysis of source code in Python programming language is considered in detail. As a practical example the analysis of the Pandas 2.2.1 project performed with the help of Svace is given. The result was the detection of 241 vulnerabilities for 590709 lines of code, which shows a high density of warnings per million lines of code and confirms the effectiveness of static analysis in ensuring software security.

Due to the use of aggressive optimizations by modern C/C++ compilers that exploit undefined behavior, there is a need for a safe compiler that does not perform such optimizations and prevents developers from using unsafe statements and expressions. Such a safe compiler based on GCC has been developed in ISP RAS, but some developers prefer Clang instead of GCC, which has mainly the same problems of exploiting undefined behavior. This paper examines the capabilities of Clang to perform safe compilation and describes the implementation of a safe compiler based on it. For the created safe compiler, the applicability in practice is shown and the impact on program performance is evaluated.

This paper presents a new version of instrumentation-based path profiling, implemented for the LCC compiler for Elbrus and SPARC processors. This profiling is intended to be used for VLIW-specific compiler optimizations, where path information and path correlations are needed. It was optimized, so profiling overhead descreased to 5.5 times on average.

The article addresses the issue of separating input information of artificial neural networks into modules using orthogonal transformations. This separation enables modular organization of neural networks with layer separation, facilitating the use of the proposed approach for distributed computing. Such an approach is required for organizing the operation of neural networks in fog and edge computing environments, as well as for high-performance computing across multiple low-performance computational nodes. The possibility of cross-layer separation of artificial neural networks using orthogonal transformations is theoretically substantiated, and practical examples of such an approach are provided. A comparison of the characteristics of modular neural networks using various types of orthogonal transformations, including the Haar wavelet transform, is conducted.

Although software development is mostly a creative process, there are many scrutiny tasks. As in other industries, there is a trend for automation of routine work. In many cases, machine learning and neural networks have become a useful assistant in that matter. Programming is not an exception: GitHub has stated that Copilot is already used to write up to 30% of code in the company. Copilot is based on Codex, a Transformer model trained on code as a sequence. However, a sequence is not a perfect representation for programming languages. In this work, we claim and demonstrate that by combining the advantages of Transformers and graph representations of code, it is possible to achieve excellent results even with comparably small models. 

The article constructs a mathematical model of distributed computing with a limited number of copies of a structured software resource; in cases of unlimited and limited parallelism by the number of processors of a multiprocessor distributed computing system, the problems of finding the minimum execution time of heterogeneous, homogeneous and identically distributed competing processes in a synchronous mode, ensuring the continuous execution of each block of software resource by all processes.

The article is devoted to the analysis of the use of machine learning algorithms to detect attacks using a custom web environment or the functionality of user applications. Learning with a teacher and clustering algorithms are considered. The dataset uses a sample of online shopping transactions collected by an e-commerce retailer. The dataset contains 39,221 transactions. To detect attacks in the web environment, the most optimal implementations of machine learning algorithms were selected after their review and comparative analysis. The most effective algorithm for detecting fraudulent transactions has been determined. We use the accuracy and running time of the algorithm as criteria. The accuracy of detecting fraudulent transactions for Random Forest, GB (Scikit-learn), GB (CatBoost) algorithms is 100%, and the KD-trees algorithm is 99,9%. The gradient boosting algorithm in the CatBoos implementation is 4,2 times faster than Random Forest, 2,4 times faster than GB Scikit-learn, 1,2 times faster than GB without using the cat_features parameter, 41,9 times faster than k-dimensional trees, 66,8 times faster than DBSCAN. The data obtained for each method is presented in the form of tables. Within the framework of this work, the parameters for evaluating the effectiveness of the algorithms under study are learning time indicators, as well as characteristics from the Confusion matrix and Classification Report for classification algorithms, and fowlkes_mallows_score, rand_score, adjusted_rand_score, Homogeneity, Completeness, V-measure for clustering algorithms.

The Residue Number System is a widely used non-positional number system. Residue Number System can be effectively used in applications and systems with a predominant proportion of addition, subtraction and multiplication operations, due to the parallel execution of operations and the absence of inter-bit carries. The reverse conversion of a number from Residue Number System to positional notation requires the use of special algorithms. The main focus of this article lies in introducing the new conversion method, which incorporates Chinese Remainder Theorem, Akushsky Core Function and rank of number. The step-by-step procedure of the conversion process is detailed, accompanied by numerical examples. The proof of the relationship between the ranks of positional characteristics using the Chinese Remainder Theorem is presented. Through careful analysis and comparison with existing transformation methods, it is concluded that the presented approach takes on average 8 % less time than the Approximate Method.

This paper presents the results of an experimental comparison of methods for the synthesis of combinational logic circuits that implement specified Boolean functions. The following methods were considered: the method of Akers, bi-decomposition, the methods of cascades, Minato-Morreale, Reed-Muller and DSD-decomposition. The comparison was based on an estimate of power, delay and area of synthesized logic circuits. The evaluation was carried out without the process of technology mapping of the circuits. These parameters were chosen because they are the main criteria for technology-independent optimization, where these methods are widely used. Boolean functions with the number of arguments from 4 to 10 were used as input data. They were generated on the basis of information on the frequency of occurrence of various NPN-equivalence classes of Boolean functions of 4 variables. As a result of the study, it was found that the Minato-Morreale method is the most universal in solving technology-independent optimization problems and can be used for different criteria.

In the era of deep learning, global-local deep neural networks are gradually replacing statistical approaches for time-series forecasting, especially for the spatiotemporal modeling field. However, the development of such methods is hindered by the lack of open benchmark datasets in this research domain. Generating synthetic data is an alternative solution to data collection, but prior works focus mainly on generating uncorrelated independent time series. In this work, we present a method for spatially correlated time-series generation. It uses a set of parametric autoregressive models for univariate time series generation in combination with the approach for sampling model parameters which allows one to simulate spatial relationships. We describe its implementation and conduct experiments showing the validity of the data for spatiotemporal modeling.

This paper proposes a method to visualize models of acyclic processes based on merging Directly-Follows Graphs (DFG) and Sankey diagrams. DFG is a popular graphical model to visualize discrete process models, while Sankey diagrams are used to represent flows of any kind. Our approach, based on flow diagrams, allows us to highlight individual cases or groups of cases in the overall model. The approach is implemented as a web-based tool that allows us, given an event log of an acyclic process, to construct and analyze the process behavior. We illustrate and evaluate the applicability of the proposed approach using learning processes as examples.

The article discusses the problem of applying runtime verification to large and complex systems such as general-purpose operating systems. When verifying the security mechanisms of operating systems, modern practices and standards require a formal security policy model (SPM). The SPM must be verified using formal model methods, and it must also be used to verify the completeness and consistency of the operating system’s security mechanisms by confirming compliance with the formal requirements of the SPM. In this case, it is convenient to have a single model suitable for both formal verification and implementation testing. For practical application, it is necessary, on the one hand, to select a subset of model language constructs suitable for both acts, and on the other hand, to develop special techniques for analyzing execution traces that allow to effectively perform thousands of test cases. The article addresses both of these issues. We present an analysis of language constructs that allow us to use the model for both verification and execution trace analysis. We also offer techniques that have been developed to optimize the runtime verification of Linux-based systems. We also implemented the proposed methods in the trace analysis tool prototype.

The paper presents the results of a numerical experiment on the identification of thermokarst objects formed as a result of climatic changes in the regions of the cryolithozone, based on satellite graphical data. An applied computer program has been developed designed to identify satellite graphical data implementing a three-layer neural network. The dependence of the object identification efficiency on various initial parameters of the neural network, such as the learning rate, the number of neurons in the hidden layer and the number of learning epochs, has been studied. The optimal values of the above parameters have been identified, providing the highest efficiency indicators of the neural network. The results obtained were compared with the data of other researchers.

In this paper was formulates the settling velocity determining problem for a single particle. To solve it, an original version of the smoothed particle method (SPH) is proposed where settling particle affects on surrounding fluid particles movement. Herewith the calculation of forces acting on particle (except for inertial forces which takes attached mass effect) is performed according by analytical mechanics formulas for a material point. This mathematical formulation and calculation algorithm was verificated by using open source code «SPH_Lab2d». Dependences for the particle settling velocity on its diameter are obtained for various cases of fluid volume discretization. The results demonstrate a good conformance this property to values, that determined by the experimental data and known phenomenological dependences for quartz sand.