Background: The goal of this work is to develop a non-invasive method in order to help detecting Alzheimer's disease in its early stages, by implementing voice analysis techniques based on machine learning algorithms. Methods: We extract temporal and acoustical voice features (e.g.Jitter and Harmonics-to-Noise Ratio) from read speech of patients in Early Stage of Alzheimer's Disease (ES-AD), with Mild Cognitive Impairment (MCI), and from a Healthy Control (HC) group. Three classification methods are used to evaluate the efficiency of these features, namely kNN, SVM and decision Tree. To assess the effectiveness of this set of features, we compare them with two sets of feature parameters that are widely used in speech and speaker recognition applications. A two-stage feature selection process is conducted to optimize classification performance. For these experiments, the data samples of HC, ES-AD and MCI groups were collected at AP-HP Broca Hospital, in Paris. Results: First, a wrapper feature selection method for each feature set is evaluated and the relevant features for each classifier are selected. By combining, By combining, for each classifier, the features selected from each initial set, we improve the classification accuracy by a relative gain of more than 30 % for all classifiers. Then the same feature selection procedure is performed anew on the combination of selected feature sets, resulting in an additional significant improvement of classification accuracy. Conclusion: The proposed method improved the classification accuracy for ES-AD, MCI and HC groups and promises the effectiveness of speech analysis and machine learning techniques to help detect pathological diseases