Identification of noisy voice data in emergency-based radio conversations using deep learning methods

Voice is the basic element used in interpersonal communication and plays a vital role especially in emergency communication. In emergency communication, the intelligibility of speech sounds is of critical importance for fast and accurate information flow. In cases where traditional communication methods fail, radio communication stands out as a reliable tool thanks to its independent structure. The biggest problem encountered in radio communication is the inability to understand noisy sounds. The use of deep learning methods, which have become prominent in recent years, provides great benefits for the correct identification of noisy sounds. Within the scope of this thesis, a new deep learning-based model that recognizes voice data has been developed to be used in emergency controls. For training and testing of the developed model, a new dataset containing six different Turkish voice data was prepared using radio recordings obtained under various environmental noise conditions. Using this dataset, six different emergency voice data were identified with 96.40% success accuracy with the developed model.