Call-sign recognition and understanding for noisy air-traffic transcripts using surveillance information

Alexander Blatt, Martin Kocour, Karel Veselý, Igor Szöke, Dietrich Klakow

Abstract

In Air Traffic Control (ATC), accurate call-sign identification is critical for safety. However, Automatic Speech Recognition (ASR) systems often struggle with the low signal-to-noise ratios of radio communications. This paper introduces a Call-sign Recognition and Understanding (CRU) system that leverages external surveillance data (ADS-B) to "denoise" transcripts. By providing the model with a list of aircraft currently in the airspace, call-sign extraction accuracy is improved by up to four times, even with high Word Error Rates (WER).

Introduction

Communication in ATC occurs over a shared radio frequency, where every command must be prefixed with a unique call-sign (e.g., "Lufthansa eight three kilo"). Automating the processing of these voice logs is challenging due to high WER in ASR systems caused by radio noise. A single misidentified digit can lead to dangerous command misattribution. This research addresses the challenge of mapping noisy transcripts like "lufthansa eight three kilo" to standardized ICAO formats like "DLH83K".

Methodology

The proposed CRU system utilizes a Transformer-based EncoderDecoderModel initialized with pretrained bert-base-uncased weights. This allows the model to leverage deep linguistic patterns while adapting to the specialized vocabulary of ATC through fine-tuning.

Contextual Integration

The core innovation is the inclusion of real-time surveillance information. The model receives a list of call-signs detected by ground-based ADS-B surveillance in the relevant airspace. This list is concatenated with the ASR transcript, providing the model with a "cheat sheet" to resolve ambiguities in the noisy text. For instance, if the ASR misses a digit but the surveillance data shows only one matching aircraft in the area, the model can intelligently fill in the blank.

Model architecture integrating transcript and surveillance data.

Data Augmentation and Training

To overcome the scarcity of labeled ATC data, the researchers developed a rule-based augmentation pipeline. This pipeline swaps call-signs, introduces real-world linguistic variations (e.g., shortening call-signs), and simulates ASR noise based on observed error distributions. Training on transcripts with simulated noise (WER 7 to 30) was found to be essential for real-world robustness, as models trained only on clean data performed poorly in practice.

Experimental Evaluation

The system was evaluated using three primary datasets:

  • Airbus Dataset: 28,000 transcripts expanded to 60,000 via augmentation for training.
  • Malorca Dataset: 1,130 human-annotated transcripts from Vienna airport, including surveillance logs.
  • LiveATC Dataset: 500 "wild" samples from Zurich, Dublin, and Atlanta airports with varying noise levels.

Results and Discussion

The integration of surveillance data yielded massive accuracy gains, multiplying the Call-sign Accuracy (CSA) by up to four times on difficult datasets. On the LiveATC dataset with a high WER of 33.1%, accuracy jumped from 11.2% (without surveillance) to 45.2% with surveillance context.

Performance metrics across varying noise levels (WER) and surveillance conditions.

The model demonstrated high resilience to "airline identifier" duplicates (e.g., multiple flights from the same airline). However, performance slightly declined when multiple aircraft shared identical call-sign numbers (e.g., DLH83K and CSA83K), particularly in high-noise scenarios. Despite this, the model successfully recovered the correct ICAO call-sign in 57% of cases where the underlying ASR transcript was only 27% correct.

Conclusion

This work demonstrates that external contextual data can significantly mitigate the impact of ASR errors in specialized domains like ATC. As part of the larger ATCO2 project, future efforts will focus on expanding entity recognition to include specific flight commands (altitudes, headings) and exploring more advanced context integration techniques.

Summary updated

BibTeX
@inproceedings{blatt2022call,
  author = {Alexander Blatt and Martin Kocour and Karel Veselý and Igor Szöke and Dietrich Klakow},
  title = {{Call-sign recognition and understanding for noisy air-traffic transcripts using surveillance information}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  doi = {10.1109/icassp43922.2022.9746301},
}
Figure