Impact of language on detecting agitation

The paper is explores the possibilities of machine learning models for agitation detection in different languages. The aim is to identify the voice features that are constant across languages and use them to develop an algorithm that can detect agitated speech by considering different languages. The research is conducted in three phases: a preparatory stage, feature extraction phase and algorithm developing phase. In the preparatory stage data sets having emotions corresponding to agitation behaviour in German and English languages were obtained. Among all voice based activities of agitation, 'rapid speech' was selected. In feature extraction stage, voice features or properties relevant to rapid speech (pitch, loudness) will be extracted. In the algorithm development phase, several machine learning models will be trained and tested to predict a level of agitation in both the languages. In-lining with the hypothesis, results indicate the difference in accuracy as we change language i.e. by using pitch and loudness as voice features in the English language 79% accuracy is achieved whereas the same features result in a 72% accuracy score for the German language. The Support Vector Classifier was most accurate in both languages.