An automated method of assessing short term memory can act as a dementia risk predictor, as poor short-term memory is strongly linked to early signs of dementia. While previous works show the feasibility of using speech to predict healthy and diagnosed dementia participants, there are still gaps in predicting 'dementia risk' and clear difficulties distinguishing early dementia with regular ageing. We extracted paralinguistic features from audio of individuals completing an over the phone episodic memory test, LOGOS. These paralinguistic features were used to discriminate between those with strong and poor short term memory performance. This work also explored various feature selection methods and tested this method across multiple datasets. Our best result was achieved using a Support Vector Machine (SVM) classifier, obtaining accuracy of 84% per audio recording.Clinical relevance- This work establishes the efficacy of using speech from older participants completing the LOGOS episodic memory test to estimate risk of dementia.