Foreign Accent Detection In Air Traffic Control Communication

Abstract

Miscommunication in the world of pilots and ATC’s is a reoccurring problem that causes terrible accidents when happening at a critical moment. Such miscommunications can arise due to the difference in nationality between the communicators due to varying pronunciations. The goal of this thesis was to create a deep neural network that would be able to determine the native language of a pilot by analysing their English speech. The languages that I have focussed on for this research are Dutch, German, and French. In order to achieve the goal, I have created a 2-dimensional Convolutional Neural Network which I fed with 1, 2, and 3 second long audio segments of pilots talking. I have tested this model on several performance measures: accuracy, loss, precision, recall and f1-score. After training the model over 20 epochs, the 1, 2, and 3 second-long segments ended up having an accuracy of 49%, 47%, and 45%. And after having performed data augmentation on the data in an attempt to reduce the overfitting problem, the 1, 2, and 3 second-long segments ended up having an accuracy of 53%, 54%, and 56%. A great improvement but still overfitting occurred. And so the results have shown that to a certain extent, it is indeed possible to predict the native accent of a pilot. For the future research this would mean that it is recommended to work on the prevention of overfitting, for example by using a much bigger dataset than the one that was used in this research.