Robust Air Traffic Control Speaker Role Classification through Combined Speaker Embeddings and Speech Understanding

Kurzfassung

Automatic speech recognition (ASR) systems in air traffic control need to distinguish between controller and pilot transmissions, a task called speaker role classification. This paper presents a robust approach by combining complementary classification methods: clustering of speaker embeddings from audio data and a rule-based classification system analyzing the semantics from generated transcripts. Our combined approach achieves 99% precision and 96% recall even on recordings from the operational environment, outperforming state-of-the-art text-based supervised learning models. Additionally, we demonstrate the robustness of our approach by evaluating the method across datasets from the tower, approach, and en-route domains, as well as analyzing performance on degraded ASR transcripts.