ATC-KD: Audio-Text Cross-modal Knowledge Distillation for Data-efficient Speech Recognition in Air Traffic Control Communications

In air traffic control (ATC), miscommunication between air traffic control officers (ATCOs) and pilots poses a significant threat to flight safety. Although recent advances in automatic speech recognition (ASR) offer promising solutions for air-ground communication processing and semantic verification, effectively integrating ASR into ATC remains challenging. This paper addresses three key issues: 1) the heterogeneity between speech and text modalities, 2) inflexible cross-modal co-learning architectures and training strategies, and 3) the scarcity of paired speech-text data in ATC communications. To overcome these challenges, we propose an Audio-Text Cross-modal Knowledge Distillation (ATC-KD) framework that efficiently leverages limited data to enhance ASR performance in ATC communications, achieving superior cross-modal semantic alignment at the encoding stage. Furthermore, we introduce a cross-modal length adapter to mitigate sequence-length disparities and a textual modality dropout strategy to reduce over-reliance on textual information. Extensive experiments under varying data conditions demonstrate the effectiveness and data efficiency of ATC-KD. Even in low-resource settings, it attains performance comparable with models trained on larger datasets, thereby mitigating data scarcity constraints.