Background:The work presented here is part of a larger study to identify voice markers for early dementia detection and it focuses on evaluating the suitability of a new approach for early diagnosis by non-invasive methods. Methods: In this context, we used class-dependent principal component analysis for feature engineering and three machine learning techniques, namely, logistic regression, support vector machines, and artificial neural networks for the automatic classification of the two classes (dementia and control subjects). Findings: We developed a non-invasive, low cost, and side-effects free approach. Our method also comprises a small number of variables and does not require heavy computing power. The developed model showed that speech parameters constitute a promising biomarker for dementia detection. Results: The obtained experimental results were satisfactory and promising when evaluated on the test set (accuracy=0.972, precision=0.983, recall=0.968, and F1-score=0.975), making the model reliable for early dementia detection.