Speech utterance classification for triage of 911 emergency calls
The utilisation of speech utterance classification for triage of 911 emergency calls is described in this research paper. Using speech utterance of caller to analyze and prioritize 911 emergency calls can reduce call wait time and enable the call taker to make better and faster decisions. The objective of the research is to examine and compare the performance of four most widely used text classification algorithms, Logistic Regression, Support Vector Machine (SVM), AdaBoost and Naive Bayes for speech utterance classification and prioritization of 911 calls. The research utilised transcripts of 911 audio calls that were created using IBM, Google and CMUSphinx Automatic Speech Recogniser (ASR) as training datasets. The classifers were trained on numerical feature vectors of bag-of-words from each training dataset and tested on manually transcribed first utterance of 911 calls and were evaluated on classification accuracy and F-score. Furthermore, these classifers were trained on three different features vectors from Bag-of-words, tf-idf weighting and word2vec model of Google ASR generated dataset to verify if word2vec features vectors improve classifier's accuracy. The result showed there was no improvement in classification accuracy when word2vec features vectors were used. Support Vector Machines achieved the highest classification accuracy of 73% when feature vectors of bag-of-words from Google ASR generated dataset were used for training.