Tvorba modelu strojového učení pro rozpoznání záměru z rádiové komunikace ATC

Thesis focuses on the development of a machine learning model for converting speech to text, specifically tailored for Czech aviation communication, achieving higher accuracy than general models. The results show that the designed machine learning model, using OpenAI's Whisper Large v3 training model, achieves much lower word error rate (WER 9%). Given the lack of freely available datasets for training, the thesis also includes the design and implementation of an extensive architecture for automatic collection, storage, and processing of radio communication in the aerodrome traffic zone at Czech airports. The work contains a detailed description of the hardware part of the architecture for reliable reception of aviation communication with minimal noise level. The software part describes and implements a scalable architecture for deploying Internet of Things devices with remote management capabilities across Czech airports, including the backend for collecting and processing a large volume of aviation recordings. Subsequently, the thesis proposes a concept verification for automatic extraction of important information such as call signs, flight paths, and pilot intentions from the transcribed recordings using large language models. For the entire project, the thesis utilizes MMSP methodology for developing the software part of the project and CRISP-DM focused on understanding, modeling, and processing data.