Muhammad Shehram Shah Syed(RMIT University), Zafi Sherhan Syed(Mehran University), Margaret Lech(RMIT University) and Elena Pirogova(RMIT University)
Dementia is a neurodegenerative disease that leads to cognitive and physical impairments. Individuals who are affected by dementia experience deterioration in their capacity to perform day-to-day tasks thereby affecting their quality of life. This paper addresses the Alzheimer's' Dementia Recognition through Spontaneous Speech (ADreSS) challenge where the objective is to propose methods for two tasks. The first task is to identify speech recordings from individuals with dementia amongst a set of recordings which also include those from healthy individuals. The second task requires participants to estimate the Mini-Mental State Examination (MMSE) score based on an individual’s speech alone. To this end, we investigated characteristics of speech paralinguistics such as prosody, voice quality, and spectra as well as VGGish based deep acoustic embedding for automated screening for dementia based on the audio modality. In addition to this, we also computed deep text embeddings for transcripts of speech. For the classification task, we achieve an accuracy of 85.42% compared to the baseline of 62.50%, meanwhile, for the regression task, we achieve an RMSE = 4.30 compared to the baseline of 6.14. These results show the promise of our proposed methods for the task of automated screening for dementia based on speech alone.