Minimum effort adaptation of automatic speech recognition system in air traffic management

Abstract

Advancements in Automatic Speech Recognition (ASR) technology is exemplified by ubiquitous voice assistants such as Siri and Alexa. Researchers have been exploring the application of ASR for Air Traffic Management (ATM) systems. Initial prototypes utilized ASR to pre-fill aircraft radar labels and achieved a technological readiness level before industrialization (TRL6). However, accurately recognizing infrequently used but highly informative domain-specific vocabulary is still an issue. This includes waypoint names specific to each airspace region and unique airline designators, e.g., “dexon” or “pobeda”. Traditionally, open-source ASR toolkits or large pre-trained models require substantial domain-specific transcribed speech data to adapt to specialized vocabularies. However, typically, a “universal” ASR engine capable of reliably recognizing a core dictionary of several hundreds of frequently used words suffices for ATM applications. The challenge lies in dynamically integrating the additional region-specific words used less frequently. These uncommon words are crucial for maintaining clear communication within the ATM environment. This paper proposes a novel approach that facilitates the dynamic integration of these new and specific word entities into the existing universal ASR system. This paves the way for “plug-and-play” customization with minimal expert intervention and eliminates the need for extensive fine-tuning of the universal ASR model. The proposed approach demonstrably improves the accuracy of these region-specific words by a factor of ≈7 (from 10% F1-score to 70%) for all rare words and ≈5 (from 13% F1-score to 64%) for waypoints.