Background: Identification of reliable, affordable and easy-to-use strategies for detection of dementia are sorely needed. Digital technologies, such as individual voice recordings, offer an attractive modality to assess cognition but methods that could automatically analyze such data without any pre-processing are not readily available. Methods: We used a subset of 1264 digital voice recordings of neuropsychological examinations administered to participants from the Framingham Heart Study (FHS), a community-based longitudinal observational study. The recordings were 73 minutes in duration, on average, and contained at least two speakers (participant and clinician). Of the total voice recordings, 483 were of participants with normal cognition (NC), 451 recordings were of participants with mild cognitive impairment (MCI), and 330 were of participants with dementia. We developed two deep learning models (a two-level long short-term memory (LSTM) network and a convolutional neural network (CNN)), which used the raw audio recordings to classify if the recording included a participant with only NC or only dementia, and also to differentiate between recordings corresponding to non-demented (NC+MCI) and demented participants. Findings: Based on 5-fold cross-validation, the LSTM model achieved a mean (±std) area under the sensitivity-specificity curve (AUC) of 0.744±0.038, mean accuracy of 0.680±0.032, mean sensitivity of 0.719±0.112, and mean specificity of 0.652±0.089 in predicting cases with dementia from those with normal cognition. The CNN model achieved a mean AUC of 0.805±0.027, mean accuracy of 0.740±0.033, mean sensitivity of 0.735±0.094, and mean specificity of 0.750±0.083 in predicting cases with only dementia from those with only NC. For the task related to classification of demented participants from non-demented ones, the LSTM model achieved a mean AUC of 0.659±0.043, mean accuracy of 0.701±0.057, mean sensitivity of 0.245±0.161 and mean specificity of 0.856±0.105. The CNN model achieved a mean AUC of 0.730±0.039, mean accuracy of 0.735±0.046, mean sensitivity of 0.443±0.113, and mean specificity of 0.840±0.076 in predicting cases with dementia from those who were not demented. Interpretation: This proof-of-concept study demonstrates the potential that raw audio recordings of neuropsychological testing performed on individuals recruited within a community cohort setting can provide a level of screening for dementia. Funding Statement: This project was supported in part by the Karen Toffler Charitable Trust, National Center for Advancing Translational Sciences, National Institutes of Health (NIH), through BU-CTSI Grant (1UL1TR001430), a Scientist Development Grant (17SDG33670323) and a Strategically Focused Research Network (SFRN) Center Grant (20SFRN35460031) from the American Heart Association, and a Hariri Research Award from the Hariri Institute for Computing and Computational Science & Engineering at Boston University, Framingham Heart Study’s National Heart, Lung and Blood Institute contract (N01-HC-25195; HHSN268201500001I) and NIH grants (R01-AG062109, R21-CA253498, R01-AG008122, R01-AG016495, R01-AG033040, R01-AG054156, R01-AG049810, U19 AG068753). Additional support was provided by Boston University's Affinity Research Collaboratives program, Boston University Alzheimer's Disease Center (P30-AG013846), the National Science Foundation under grants DMS-1664644, CNS-1645681, and IIS-1914792, the Office of Naval Research under grant N00014-19-1-2571, and the NIH under grant R01-GM135930. Declaration of Interests: None to declare. Ethics Approval Statement: The study was approved by the Institutional Review Boards of Boston University Medical Center and all participants provided written consent.