Accent Based Transformer Ensemble for Air Traffic Control Speech Recognition

Air traffic controllers (ATC) communicate with pilots via speech, using radio channels. The noise and the diverse accents in the speech make it challenging to understand, war-ranting the need for a robust transcription system. There is little work regarding accent-based ensembles to tackle ATC automatic speech recognition. Moreover, there is limited research focusing on Automatic Speech Recognition (ASR) for Indian-accented english ATC speech. We present a novel accent based end-to-end transformer ensemble model to transcribe the communications between ATC and pilots. The proposed methodology outperforms an existing ensemble method (the confidence-based ensemble) by 10.5% in accuracy. Furthermore, it achieves an improvement of 3% over the expected accuracy calculated using the ‘Whisper’ ASR model’s performance across different parameter sizes.