There is a risk that voice messages from in-vehicle information systems may cause a driver to be distracted while driving. To avoid such a risk the message systems need to be adapted to drivers' mental workload. Such adaptive systems deliver voice messages when drivers' mental workload is low and postpone the messages when the driver workload is high. It is important for the system to estimate the current driver workload from car sensors such as car speed, steering wheel angle, accelerator pedal position and so on. In order to find some relations between the driver's mental workload and the data from car sensors, a dual task experiment was conducted on a public road. In the experiment, participants performed a memory task while driving an experimental car. At the same time, the data from the car sensors were recorded. The correlation coefficients between the memory task performance and the data from car sensors showed that release of the accelerator pedal was the most significant indicator of workload. Based on these results a workload estimator was developed, which has been applied to a voice information delivery test system. The potential of a voice information system that adapts to the driver's mental workload was evaluated.