Radiotelephony remains the primary medium for pilot-controller communication, yet extracting structured information from spoken exchanges is challenging. Deep learning approaches often depend on large annotated datasets, limiting use in data-scarce environments. This study evaluates open-source Large Language Models for Structured Information Extraction from ATC communications, with applications in assisting or automating pseudo-pilot tasks. We evaluate Llama 3.3 (70B) with baseline prompting and Gemma 3 (4B) with baseline and fine-tuned variants on 496 utterances from NLR’s ATM simulator: NARSIM (NLR ATC real-time simulator). Performance is assessed on human transcripts and ASR outputs from Whisper models, with varying prompt contexts. Cross-sector generalization is tested across two ATC sectors. Using manual scoring, Llama 3.3 achieves micro-F1 0.95 on human transcripts and 0.86 on fine-tuned Whisper outputs. While Gemma 3 performed weaker in its baseline form, fine-tuning on a small sample led to notable improvements. Results demonstrate the potential of LLMs for ATC applications without the need for large annotated datasets. ... Read more

Current advancements in machine learning have provided new architectures, such as encoder-decoder transformers, for automatic speech recognition. For generic speech recognition, very high accuracies are already achievable. However, in air traffic control, automatic speech recognition models traditionally rely on domain-specific models constructed from limited training data. This study introduces this newly developed transformer model for air traffic control and provides a set of fully open automatic speech recognition models with high accuracies. This paper demonstrates how a large-scale, weakly supervised automatic speech recognition model, Whisper, is fine-tuned with various air traffic control datasets to improve model performance. We also evaluated the performance of different sizes of Whisper models. In the end, it was possible to achieve word error rates of 13.5% on the ATCO2 dataset and 1.17% on the ATCOSIM dataset with a random split (or 3.88% with speaker split). The study also reveals that finetuning with region-specific data can enhance performance by up to 60% in real-world scenarios. Finally, we have open-sourced the code base and the models for future research. ... Read more