This paper describes research behind a Large-Vocabulary Continuous Speech Recognition (LVCSR) system for the transcription of Senate speeches for the Polish language. The system utilizes severalcomponents: a phonetic transcription system, language and acoustic model training systems, a Voice Activity Detector (VAD), a LVCSR decoder, and a subtitle generator and presentation system. Some of the modules relied on already available tools and some had to be made from the beginning but the authors ensured that they used the most advanced techniques they had available at the time. Finally, several experiments were performed to compare the performance of both more modern and more conventional technologies.

The article considers information technology for the realization of human communication using residual human capabilities, obtained by organizing text entry using mobile and auxiliary devices. The components of the proposed technology are described in detail: the method for entering text information to realize the possibility of introducing a limited number of controls and the method of predicting words that are most often encountered after words already entered in the sentence. A generalized representation of the process of entering text is described with the aid of an ambiguous virtual keyboard and the representation of control signals for the selection of control elements. The approaches to finding the optimal distribution of the set of alphabet characters for different numbers of control signals are given. The method of word prediction is generalized and improved, the statistical language model with "back-off" is used, and the approach to the formation of the training corpus of the spoken Ukrainian language is proposed.

This paper explores cost-effective alternatives for resource-constrained environments in the context of language models by investigating methods such as quantization and CPUbased model implementations. The study addresses the computational efficiency of language models during inference and the development of infrastructure for text document processing. The paper discusses related technologies, the CLARIN-PL infrastructure architecture, and implementations of small and large language models. The emphasis is on model formats, data precision, and runtime environments (GPU and CPU). It identifies optimal solutions through extensive experimentation. In addition, the paper advocates for a more comprehensive performance evaluation approach. Instead of reporting only average token throughput, it suggests considering the curve’s shape, which can vary from constant to monotonically increasing or decreasing functions. Evaluating token throughput at various curve points, especially for different output token counts, provides a more informative perspective.