Multi-part Analysis and Techniques for Air Traffic Speech Recognition

The general English speech recognition is based on the techniques of n-grams where the words before and after are predicted and the utterance prediction is produced. At the same time, having a significantly lengthier n-gram has its own impact in training and the accuracy. Shorter n-grams require the utterances to be split and predicted than using the complete utterance. This article discusses specific techniques to address the specific problems in Air Traffic Speech, which is a medium length utterance domain. Moving from the adapted language models (LMs) to rescored LM, a combined technique of syntax analysis along with a deep learning model is proposed, which improves the overall accuracy. It is explained that this technique can help to adapt the proposed method for different contexts within the same domain and can be successful.