Reducing the latency and model size has always been a significant research problem for live Automatic Speech Recognition (ASR) application scenarios. Along this direction, quantization become an increasingly popular approach to compress neural networks reduce computation cost. Most of existing practical ASR systems apply post-training 8-bit quantization. To achieve higher compression rate without introducing additional performance regression, in study, we propose develop 4-bit models with...

Based on the analysis of main normative documents defining directions ensuring national security and development Russia, authors substantiate need to form a strategic program target document in this area, integrating tasks implementing demographic family policy State. The provide their understanding content security, which consider as one components Russian Federation. conclusion that goal context modern challenges is improve indicators through reproduction indigenous population reflecting...

In this paper, we investigate novel quantization approaches to reduce memory and computational footprint of deep neural network (DNN) based keyword spotters (KWS). We propose a new method for KWS offline online quantization, which call dynamic where quantize DNN weight matrices column-wise, using each column's exact individual min-max range, the layers' inputs outputs are quantized every input audio frame individually, range output vector. further apply quantization-aware training approach...

Model fine-tuning and adaptation have become a common approach for model specialization downstream tasks or domains. Fine-tuning the entire subset of parameters using light-weight has shown considerable success across different tasks. large number domains typically requires starting new training job every domain posing scaling limitations. Once these models are trained, deploying them also poses significant scalability challenges inference real-time applications. In this paper, building upon...

Based on the analysis of main risks information environment for minors, authors substantiate need to form a constitutional and legal model ensuring security minors. The represent their understanding concepts “threat,” “challenge,” “danger,” which are close in categorical series with concept “risk,” is presented, propose definitions “information security,” “security worldview.” goals minors both creation maintenance most favorable conditions adaptation minor modern society, contributes his...

Simultaneous speech-to-speech translation (S2ST) holds the promise of breaking down communication barriers and enabling fluid conversations across languages. However, achieving accurate, real-time through mobile devices remains a major challenge. We introduce SimulTron, novel S2ST architecture designed to tackle this task. SimulTron is lightweight direct model that uses strengths Translatotron framework while incorporating key modifications for streaming operation, an adjustable fixed delay....

Large-scale universal speech models (USM) are already used in production. However, as the model size grows, serving cost grows too. Serving of large is dominated by that why reduction an important research topic. In this work we focused on using weights only quantization. We present binarization USM Recurrent Neural Network Transducer (RNN-T) and show its can be reduced 15.9x times at word error rate (WER) increase 1.9% comparison to float32 model. It makes it attractive for practical applications.

.

Large-scale universal speech models (USM) are already used in production. However, as the model size grows, serving cost grows too. Serving of large is dominated by that why reduction an important research topic. In this work we focused on using weights only quantization. We present binarization USM Recurrent Neural Network Transducer (RNN-T) and show its can be reduced 15.9x times at word error rate (WER) increase 1.9% comparison to float32 model. It makes it attractive for practical applications.

Introduction. Technosocial transformations resulting from the industrial and digital revolution Industry 4.0 initiate multiple changes in organization of social life, culture, law, ways thinking, activities, values, worldview. There is reason to believe that there a process formation new kind which not subspecies or part information legal culture. Theoretical Basis. Methods. Information, technological, have become general factors are localized at level tools, auxiliary means organizing human...

Training stability of large language models(LLMs) is an important research topic. Reproducing training instabilities can be costly, so we use a small model with 830M parameters and experiment higher learning rates to force models diverge. One the sources instability growth logits in attention layers. We extend focus previous work look not only at magnitude but all outputs linear layers Transformer block. observe that high rate L2 norm layer grow each step diverges. Specifically QKV, Proj FC2...

Purpose. The article substantiates the need to introduce concept of “legal information culture” into scientific circulation, formulate a definition this concept, its elemental composition and practical significance in context digital transformation society state. Legal culture carries humanistic-value multidimensional potential accumulated cultural legal experience mankind, which is relevant modern conditions. It on basis that it possible form model new type: culture. Methodology:...

The author considers the issues of legal information significance as most important and necessary requirement for formation implementation modern Russian policy, analyzes problems content classification information, provides own definitions policy taking into account update various aspects its understanding, reveals dependence efficiency from communication space development level, differentiates between shows that variety forms (law-making, enforcement, doctrinal etc.) makes actual...