This paper presents a user verification system on mobile phones that is based on keystroke dynamics derived from a touchable keyboard. The touch keystroke dynamics dataset are collected using a developed mobile application in which, unlike other systems, no specific text is required. Two scenarios were considered: few-training and more-training datasets. The Median Vector Proximity classifier is applied on both datasets and the performance of the system is investigated using a different number of features. Using few-training dataset, the average EER were 12.9% and 12.2% for 31 and 33 features respectively. Using more- training dataset brings improved results with EER=0.76% and EER=0.39% for 31 and 33 features respectively. The Medians Vector Proximity becomes more accurate when increasing the training data. Also, using more features reduced the average EER by 0.7% and 0.37% in few-training and more-training datasets respectively. The proposed system is compared against other systems and shows promising results.