In recent years, as the development of Mobility On Demand (MOD) services, smartphone-based car-sharing applications became more and more popular. The accurate estimation on waiting-time for a reserved car is a significant aspect to evaluate the car-sharing applications. Waiting-time prediction can be achieved by estimating the traffic situation from the probe information, such as speed and position, which are provided by a huge number of shared cars. Thus, developing a smartphone-based vehicle speed estimation system could improve the performance of the MOD services, and an independent smartphone-based application without any connection with the car also satisfies the convenience demand of MOD service applications. In order to achieve this aim, we proposed the DeepSpeedometer system, which estimates vehicle speed from the accelerometer and gyroscope in the smartphone using a two-layer Long Short-Term Memory (LSTM). For evaluating the performance of the proposed system, we performed a series of experiments, including the analysis on the rationality of the proposed algorithm itself and the comparison with other methods as well. The experimental results show the robustness of our system in different traffic scenarios.