Automatic Speech Recognition Benchmark for Air-Traffic Communications

Automatic Speech Recognition Benchmark for Air-Traffic Communications

Stať ve sborníku v databázi WoS či Scopus

Advances in Automatic Speech Recognition (ASR) over the lastdecade opened new areas of speech-based automation such as inAir-Traffic Control (ATC) environments. Currently, voice communicationand data links communications are the only wayof contact between pilots and Air-Traffic Controllers (ATCo),where the former is the most widely used and the latter is anon-spoken method mandatory for oceanic messages and limitedfor some domestic issues. ASR systems on ATCo environmentsinherit increasing complexity due to accents from non-English speakers, cockpit noise, speaker-dependent biases andsmall in-domain ATC databases for training. Hereby, we introduceCleanSky EC-H2020 ATCO2, a project that aims todevelop an ASR-based platform to collect, organize and automaticallypre-process ATCo speech-data from air space. Thispaper conveys an exploratory benchmark of several state-ofthe-art ASR models trained on more than 170 hours of ATCospeech-data. We demonstrate that the cross-accent flaws dueto speakers accents are minimized due to the amount of data,making the system feasible for ATC environments. The developedASR system achieves an averaged word error rate (WER)of 7.75% across four databases. An additional 35% relative improvementin WER is achieved on one test set when training aTDNNF system with byte-pair encoding.

Advances in Automatic Speech Recognition (ASR) over the lastdecade opened new areas of speech-based automation such as inAir-Traffic Control (ATC) environments. Currently, voice communicationand data links communications are the only wayof contact between pilots and Air-Traffic Controllers (ATCo),where the former is the most widely used and the latter is anon-spoken method mandatory for oceanic messages and limitedfor some domestic issues. ASR systems on ATCo environmentsinherit increasing complexity due to accents from non-English speakers, cockpit noise, speaker-dependent biases andsmall in-domain ATC databases for training. Hereby, we introduceCleanSky EC-H2020 ATCO2, a project that aims todevelop an ASR-based platform to collect, organize and automaticallypre-process ATCo speech-data from air space. Thispaper conveys an exploratory benchmark of several state-ofthe-art ASR models trained on more than 170 hours of ATCospeech-data. We demonstrate that the cross-accent flaws dueto speakers accents are minimized due to the amount of data,making the system feasible for ATC environments. The developedASR system achieves an averaged word error rate (WER)of 7.75% across four databases. An additional 35% relative improvementin WER is achieved on one test set when training aTDNNF system with byte-pair encoding.

Speech Recognition, Air Traffic Control, TransferLearning, Deep Neural Networks, Lattice-Free MMI

Speech Recognition, Air Traffic Control, TransferLearning, Deep Neural Networks, Lattice-Free MMI

ZULUAGA-GOMEZ, J.; MOTLÍČEK, P.; ZHAN, Q.; VESELÝ, K.; BRAUN, R.

2021

25.10.2020

International Speech Communication Association

Shanghai

Proceedings of Interspeech 2020

1990-9772

Proceedings of Interspeech

2020

10

Francouzská republika

2297

2301

5

https://isca-speech.org/archive/Interspeech_2020/pdfs/2173.pdf

zuluaga-gomez_Interspeech2020_2173