ROSE Simulation acquires a test license for the SELF-MADE-ATC prototype

ROSE Simulation and DLR met on April 23rd and 24th 2026 in Eschau at the ROSE facilities for a feedback session on the first version of the SELF-MADE-ATC prototype. The DLR SELF-MADE-ATC project enables air traffic control organisations and ATM system manufacturers to maintain and adapt an automatic speech recognition and understanding system (ASRU) with small effort using automation and artificial intelligence. DLR has developed its first SELF-MADE-ATC prototype.

Workshop participants from ROSE Simulation and DLR, in the ROSE Simulation’s tower simulator  in April’26

ROSE Simulation provides cutting-edge technology and customized solutions for air traffic control (ATC) training environments. ROSE Simulation is the Creator of the AIR Suite, a fully-integrated set of ATC training solutions, composed of the AMOS Tower Simulator, the IRIS VoiceCom System, the ROSE Radar Simulator and SARA, the newest ROSE product reducing the work of simulation pilots with speech recognition and understanding support.

For the SELF-MADE-ATC project, ROSE has recorded more than 2000 controller utterances in the simulation environment of EPN simulating Copenhagen approach. During the feedback session in Eschau, the recorded utterances were split into training and test data. A small test dataset was manually transcribed by DLR. The SELF-MADE-ATC prototype automatically transcribes the recorded utterance using different Speech-to-Text models. The resulting transcriptions are then fed to a large language model (LLM) to generate a new pseudo-label (or a single improved transcription), which is subsequently used to fine-tune a baseline speech recognition model. The baseline model was never trained with utterances from Copenhagen approach. The final fine-tuned model outperformed the baseline model, achieving a relative reduction in the word error rate of more than 50%.

This first prototype concentrates on fine-tuning speech recognition models by leveraging large unlabelled ATC voice recording datasets. The next prototype version will also support updating the speech understanding building blocks, which will enable the support of new phraseology, new airline designators or waypoint names without additional manual transcription.

Scroll to Top