Methods and systems for training a computer to recognize and verify an individual's signature are provided. One illustrative method extracts a plurality of both global and local features from a relatively small sample of handwriting samples. In one such embodiment, only 5 samples are needed from an individual without requiring forgeries. In yet another embodiment, only three global parameters are utilized, thus reducing the complexity, and processing power, of the system. Utilizing such few global parameters also facilitates fine tuning of the systems. Further aspects of the invention relate to a multi-stage statistical system for on-line signature verification. In one embodiment, the system may comprise a simplified GMM model built on global signature properties and a left-to-right HMM model based on segmental features. In one embodiment, specific strategies are utilized to create model simplification and initialization in contrast to general GMM and HMM models.