Minimizing Errors in Air Traffic Speech Using Rule Based Algorithms

Abstract

The application of automatic speech recognition for air traffic speech has attracted a lot of attention in recent times. The application of machine learning and deep learning models can address the challenges of speech recognition to some extent. This work focuses on understanding the intricacies of domain characteristics and how to leverage them to improve speech recognition.This work focuses on applying rule-based techniques in the postprocessing phase of speech recognition. The post-processing stage is defined as
a stage where a set of algorithms are applied to the decoded output of speech recognition. Multiple techniques, like syntax separation techniques, co-occurrence-based clustering techniques, and string distance algorithms, are discussed in this work. The choice of these algorithms is based on an understanding of the significance of air traffic speech domain characteristics. Machine learning and deep learning models were applied in feature selection, language model generation, or acoustic model generation. Approaches based on rules could be selectively applied as an incremental update to language models or acoustic model generation. This work was able to improve the accuracy by 9% by applying selective algorithms for error detection and correction. Comparisons of different techniques were discussed when multiple techniques were used for clustering and string distance calculations. One of the good observations of this work is that leveraging the characteristics of speech in this domain helped improve accuracy. The improvement in accuracy was
seen in two scenarios. One scenario has a complete utterance, and the other scenario has syntax-separated utterances. A rule-based algorithm was proposed for syntax separation.