A Noise-robust U2 Scheme for Automatic Speech Recognition of Air Traffic Voice Communication

Air traffic voice communication (ATC Communication) involves the transmission of operational commands between controllers and pilots through voice. This communication is crucial for maintaining the safety and order of air traffic management. However, due to manual execution, errors, omissions and errors often occur, leading to safety accidents. Automatic speech recognition (ASR) solution tailored to the characteristics of air traffic speech is crucial to achieving real-time recognition, reducing ambiguity and preventing errors. This paper presents a Noise-Robust U2 scheme for ATC communication ASR. The scheme enhances the speech under severe noise interference and achieves high efficiency and accuracy while ensuring the robustness of the recognition process. The scheme includes two key components: a noise-robust speech enhancement model based on the two-layer gating cycle unit (GRU), which extracts compact features through the Bark frequency inverted spectrum coefficient (BFCC) and captures short-and long-term dependencies through the GRU; the second is the U2 ASR method, which provides real-time flow identification and optimized output combined with Transformer encoder and Connectionist temporal classification(CTC) decoder and attention-based(AED) decoder. Finally, noise data from Noise-92, clean voice data from Aishell-1 and ATC communication recording data collected by an airport in China were selected to verify the feasibility of the proposed scheme.