Telephone CPR Instructions in Cardiac Arrest

Abstract

This thesis focus on developing a dataset of recordings between a caller and a dispatcher from Emergency Communication Centre during situations involving cardiac arrest. It also focused on developing and implementing a speech recognition system that would analyze 10 keywords in the dataset. One of the methods tested was to design a convolutional neural network to identify the small 1 second keywords. A separate dataset with these keywords were developed and model was trained and validated. A speech segmentation algorithm was developed to identify the boundaries of the words in the dataset. The algorithm analyzed the different energy levels in the signal to separate words from silence and to find the word boundaries. These words where then classified with the CNN model. If the likelihood of the word belonging to one of the classes was less then 99.99 %, then the word was not classified. However, if the likelihood was more then 99.99 % the word was classified and the output was written as text. The other method was to nd the keywords using an existing speech recognition system. Google API speech recognition system was used to transcribe the recordings to text. Both the transcriptions from both of the methods were compared with the real transcription. The results gave a word error rate of 76.4% and false alarm rate of 282% for the CNN model and 32.48 % and 1.91% for the Google model. In conclusion the Google speech recognition model was better at transcribing the dataset then the developed CNN model.