While beam search improves speech recognition quality over greedy decoding, standard implementations are slow, often sequential, and CPU-bound. To fully leverage modern hardware capabilities, we present a novel open-source FlexCTC toolkit for fully GPU-based beam decoding, designed for Connectionist Temporal Classification (CTC) models. Developed entirely in Python and PyTorch, it offers a fast, user-friendly, and extensible alternative to traditional C++, CUDA, or WFST-based decoders. The toolkit features a high-performance, fully batched GPU implementation with eliminated CPU-GPU synchronization and minimized kernel launch overhead via CUDA Graphs. It also supports advanced contextualization techniques, including GPU-powered N-gram language model fusion and phrase-level boosting. These features enable accurate and efficient decoding, making them suitable for both research and production use.

Despite Arabic being one of the most widely spoken languages, the development of Arabic Automatic Speech Recognition (ASR) systems faces significant challenges due to the language’s complexity, and only a limited number of public Arabic ASR models exist. While much of the focus has been on Modern Standard Arabic (MSA), there is considerably less attention given to the variations within the language. This paper introduces a universal methodology for Arabic speech and text processing designed to address unique challenges of the language. Using this methodology, we train two novel models based on the FastConformer architecture: one designed specifically for MSA and the other, the first unified public model for both MSA and Classical Arabic (CA). The MSA model sets a new benchmark with state-of-the-art (SOTA) performance on related datasets, while the unified model achieves SOTA accuracy with diacritics for CA while maintaining strong performance for MSA. To promote reproducibility, we open-source the models and their training recipes.

Traditional automatic speech recognition (ASR) models output lower-cased words without punctuation marks, which reduces readability and necessitates a subsequent text processing model to convert ASR transcripts into a proper format. Simultaneously, the development of end-to-end ASR models capable of predicting punctuation and capitalization presents several challenges, primarily due to limited data availability and shortcomings in the existing evaluation methods, such as inadequate assessment of punctuation prediction. In this paper, we introduce a LibriSpeech-PC benchmark designed to assess the punctuation and capitalization prediction capabilities of end-to-end ASR models. The benchmark includes a LibriSpeech-PC dataset with restored punctuation and capitalization, a novel evaluation metric called Punctuation Error Rate (PER) that focuses on punctuation marks, and initial baseline models. All code, data, and models are publicly available.

This paper introduces a novel Russian speech dataset called Golos, a large corpus suitable for speech research. The dataset mainly consists of recorded audio files manually annotated on the crowd-sourcing platform. The total duration of the audio is about 1240 hours. We have made the corpus freely available to download, along with the acoustic model with CTC loss prepared on this corpus. Additionally, transfer learning was applied to improve the performance of the acoustic model. In order to evaluate the quality of the dataset with the beam-search algorithm, we have built a 3-gram language model on the open Common Crawl dataset. The total word error rate (WER) metrics turned out to be about 3.3% and 11.5%.

Clustering is a well-known and important problem with numerous applications. The graph-based model is one of the typical cluster models. In the graph model generally clusters are defined as cliques. However, such approach might be too restrictive as in some applications, not all objects from the same cluster must be connected. That is why different types of cliques relaxations often considered as clusters. In our work, we consider a problem of partitioning graph into clusters and a problem of isolating cluster of a special type where by cluster we mean highly connected subgraph. Initially, such clusterization was proposed by Hartuv and Shamir. And their HCS clustering algorithm was extensively applied in practice. It was used to cluster cDNA fingerprints, to find complexes in protein-protein interaction data, to group protein sequences hierarchically into superfamily and family clusters, to find families of regulatory RNA structures. The HCS algorithm partitions graph in highly connected subgraphs. However, it is achieved by deletion of not necessarily the minimum number of edges. In our work, we try to minimize the number of edge deletions. We consider problems from the parameterized point of view where the main parameter is a number of allowed edge deletions. The presented algorithms significantly improve previous known running times for the Highly Connected Deletion (improved from \cOs\left(81^k\right) to \cOs\left(3^k\right)), Isolated Highly Connected Subgraph (from \cOs(4^k) to \cOs\left(k^{\cO\left(k^{\sfrac{2}{3}}\right)}\right) ), Seeded Highly Connected Edge Deletion (from \cOs\left(16^{k^{\sfrac{3}{4}}}\right) to \cOs\left(k^{\sqrt{k}}\right)) problems. Furthermore, we present a subexponential algorithm for Highly Connected Deletion problem if the number of clusters is bounded. Overall our work contains three subexponential algorithms which is unusual as very recently there were known very few problems admitting subexponential algorithms.

Improving user experience through personalized recommendations is crucial to organizing the abundance of data on news websites. Modeling user preferences based on implicit feedback has recently gained lots of attention, partly due to growing volume of web generated click stream data. Matrix factorization learned with stochastic gradient descent has successfully been adopted to approximate various ranking objectives. The aim of this paper is to test the performance of learning to rank approaches on the real-world dataset and apply some simple heuristics to consider temporal dynamics present in news domain. Our model is based on WARP loss with changes to classic factorization model.

The Secluded Path problem introduced by Chechik et al. in [ESA 2013] models a situation where a sensitive information has to be transmitted between a pair of nodes along a path in a network. The measure of the quality of a selected path is its exposure, which is the total weight of vertices in its closed neighborhood. In order to minimize the risk of intercepting the information, we are interested in selecting a secluded path, i.e. a path with a small exposure. Similarly, the Secluded Steiner Tree problem is to find a tree in a graph connecting a given set of terminals such that the exposure of the tree is minimized. In this work, we obtain the following results about parameterized complexity of secluded connectivity problems. We start from an observation that being parameterized by the size of the exposure, the problem is fixed-parameter tractable (FPT). More precisely, we give an algorithm deciding if a graph G with a given cost function w:V(G)->N contains a secluded path of exposure at most k with the cost at most C in time O(3^{k/3}(n+m) log W), where W is the maximum value of w on an input graph G. Similarly, Secluded Steiner Tree is solvable in time O(2^{k}k^2 (n+m) log W). The main result of this paper is about "above guarantee" parameterizations for secluded problems. We show that Secluded Steiner Tree is FPT being parameterized by r+p, where p is the number of the terminals, l the size of an optimum Steiner tree, and r=k-l. We complement this result by showing that the problem is co-W[1]-hard when parameterized by r only. We also investigate Secluded Steiner Tree from kernelization perspective and provide several lower and upper bounds when parameters are the treewidth, the size of a vertex cover, maximum vertex degree and the solution size. Finally, we refine the algorithmic result of Chechik et al. by improving the exponential dependence from the treewidth of the input graph.