Iterative Learning of Speech Recognition Models for Air Traffic Control

Abstract

Automatic Speech Recognition (ASR) has recently proved to be a useful tool to reduce the workload of air traffic controllers leading to significant gains in operational efficiency. Air Traffic Control (ATC) systems in operation rooms around the world generate large amounts of untranscribed speech and radar data each day, which can be utilized to build and improve ASR models. In this paper, we propose an iterative approach that utilizes increasing amounts of untranscribed data to incrementally build the necessary ASR models for an ATC operational area. Our approach uses a semi-supervised learning framework to combine speech and radar data to iteratively update the acoustic model, language model and command prediction model (i.e. prediction of possible commands from radar data for a given air traffic situation) of an ASR system. Starting with seed models built with a limited amount of manually transcribed data, we simulate an operational scenario to adapt and improve the models through semi-supervised learning. Experiments on two independent ATC areas (Vienna and Prague) demonstrate the utility of our proposed methodology that can scale to operational environments with minimal manual effort for learning and adaptation.

On the Learning Dynamics of Semi-Supervised Training for ASR

“…[11,12] -that initial system quality directly affects subsequent SST success, we here hypothesize that optimally-performed SST may specifically be much more sensitive to the quality of the LM than the AM or system as a whole: for a sufficient LM may be able to counteract the weaknesses of a poor AM (hence poor frame-level predictions), allowing for SST to make gains in situations otherwise lost to error-propagation (c.f. [13]). By establishing our own semi-supervised pipeline and comparing how varyingquality seed AMs, varying-quality LMs, and, importantly, combinations of such AMs and LMs, affect final Word Error Rate (WER) achieved, we explore how seed acoustic model (AM) quality and, independently, language model (LM) quality impact upon SST success.…”

ATCSpeechNet: A multilingual end-to-end speech recognition framework for air traffic control systems

“…Srinivasamurthy et al studied the ASR approach [52] in the ATC domain, in which the extracted prior knowledge is applied to improve the final performance. A semi-supervised approach was developed to achieve the iterative training of the ASR model [53].…”

Call-Sign Recognition and Understanding for Noisy Air-Traffic Transcripts Using Surveillance Information

“…New systems like controller-pilot data link communications (CPDLC), which use text based communication, reduce the load on the voice communication channels [1]. Projects like AcListant and MALORCA 1 aim to support the ATCO by speech recognition systems [2,3]. The problem with developing such systems, is the lack of training…”