Error Correction of ASR in Air Traffic Control

Automatic Speech Recognition (ASR) for Air Traffic Control (ATC) is designed to convert ATC voice commands into text for Pilot-Controller Voice Communications (PCVCs) and other downstream tasks. Proper recognition of ATC instructions is critical to flight efficiency and flight safety. In order to improve the performance of speech recognition in air traffic control, we propose a transcription-based error correction method exploiting parallel Sequence-to-Sequence (S2S) models of Conformer and Transformer, which translate ASR output into corrected text. In this design, the models are trained by paired data of ASR hypothesis and ASR reference. To acquire ASR transcriptions with recognition errors, we perform data augmentation by applying time-frequency masking, adding various noises to the ATC speech, using n-best outputs of the ASR decoder, and adding correct text to the training data. The method also sets similarity threshold to ensure the correctability of the input and performs the output-checking to obtain reliable output. Experiments show that the proposed method reduces Character Error Rate (CER) and Sentence Error Rate (SER) by up to 1% and 8.3%, respectively.