Gears are the main components of power transmissions and are subjected to high cyclic load regime which can lead to premature fracture of the gear teeth. In order to prevent such events, research on gear condition monitoring and fault diagnostics techniques have received considerable attention. Machine learning (ML) applications have been widely combined with vibration measurement and analysis techniques for fault diagnostics in gearboxes and the majority of current techniques rely on experiments to generate training data. Despite the recognized advantages of using simulated data to train ML classifiers, this approach is still not a widespread practice. This paper proposes a simulation-driven ML framework to estimate the crack size in a gear tooth using simulated vibrations signals. Firstly, a 6-degrees-of-freedom model of a one-stage gearbox was developed to simulate the dynamic behavior of a cracked pinion. Secondly, a sample with 900 simulated vibration signals was generated considering 4 different crack sizes in the pinion tooth. Thirdly, the features of the vibration signals were extracted using 20 statistical indicators and, then, the extracted features were used to train and test 4 machine learning classifiers. Several performance evaluation metrics were computed, and the performance of the ML classifiers was compared and discussed. It was shown that the Decision Tree Classifier (DTC) performed best in the identification of small crack sizes regardless of the random selection of the training subsample.