FEASIBILITY OF AUTOMATED VOICE-BASED COGNITIVE ASSESSMENT ON A CONSUMER VOICE PLATFORM

Abstract

Speech is a complex and uniquely human behaviour including both motor and cognitive components. Neurological and neurodegenerative diseases impact on different aspects of speech and language. Speech is a response modality in a large number of established cognitive tests tapping non-language aspects of cognition including episodic memory, working memory and executive function. In clinical studies, accuracy and consistency in scoring and administration are still problematic; computerised cognitive batteries such as CANTAB can eliminate this source of error by automating test administration and scoring. Computerising verbal tests has not been feasible to date due to scoring errors introduced by the relatively poor performance of speech recognition technology. However, state-of-the-art speech recognition incorporating deep learning and other recent advances have prompted us to revisit prior feasibility assumptions. We developed a short voice-based cognitive battery with tasks assessing episodic memory, working memory and executive function using off-the-shelf speech recognition technology. The battery was composed of automated versions of the Digit Span, Auditory-Verbal Learning and Verbal Paired Associates tests. We assessed twenty participants using the automated battery and simultaneous independent audio. Participants were also assessed using a standard CANTAB battery measuring similar cognitive domains. Analysis focused on assessing 1) the reliability of the scoring 2) the construct validity of the battery compared to other cognitive tests. Expected patterns of covariance were found between the standard CANTAB battery and the automated voice battery. The performance of the speech recognition software varied between tasks, with digit recognition performance during the digit span task superior to unconstrained word recognition during the paired associates learning task. This corresponds to the intrinsic difficulty of the recognition task. Automated voice-based assessment of cognition may now be a practical and accurate means of administering cognitive tests remotely, for example in population screening or home based monitoring. Further assessment of the feasibility and effectiveness of the automated voice battery is required.