Samodejna priprava prometnih poročil za obveščanje javnosti prek radijskih postaj

In this paper, we present the development of a system for automatic generation of traffic reports for radio reporting using large-scale language models. The system is based on data from the national traffic portal NAP, from which data on traffic events and road works is regularly obtained. The obtained data is stored, normalized and prepared for further classification and text generation. Special emphasis is placed on comparing different approaches to the use of language models, namely cloud-based models and locally learned or adapted models. The comparison is carried out for the tasks of traffic event classification and traffic report generation, taking into account the quality of the outputs, response speed, usage costs, robustness and suitability for the Slovenian language. To generate regular traffic reports, we use a two-stage approach, in which traffic records are first converted into structured semantic fragments, and then into the final Slovenian text suitable for radio reading. The paper discusses the data pipeline, input data processing, model comparison and the role of humans in the final editorial judgment.