Detecting agitated speech : A neural network approach

Hetterscheid, K.J.T. (2020)

Agitation is common across neuropsychiatric disorders such as dementia. It is considered as a symptom of distress which contributes to disability, institutionalization, and diminished quality of life for patients and their caregivers. Literature suggests that agitation can be monitored or detected by abnormal vocal and physical activities. For the scope of this research, voice-based activities were used.The aim is to construct a suitable neural network that can classify these voice activities. Several datasets (RAVDESS,TESS and ElderReact) are used to train and test a Recurrent Neural Network (RNN). This network is build using Long Short Term Memory(LSTM)layers and Bidirectional LSTM layers. Several combinations of these are used and compared to find the most suitable combination.The proposed model reaches an accuracy of 86%, which is in line with other state-of-the-art approaches.
Hetterscheid-BA-EEMCS.pdf