The excellent sensing performance and privacy-preserving nature of millimeter-wave (mmWave) radar has led to its increasing prominence in the field of human sensing, enabling it to support a wide range of human sensing applications. However, the scarcity of mmWave datasets and the high cost of data collection limit the capability of deep learning. In this paper, we propose PmSyn, a system that synthesizes mmWave signals based on human body model provided by public human motion database and trains neural networks with signatures extracted from synthesized signals to achieve activity recognition and skeleton tracking. We analyze in depth the reflection of the human body model and the propagation principle of the signal. We propose to inject real user’s shape information into the public human body model to endow the synthesized signals with personalization. Experiments show that for activity recognition, PmSyn achieves 93.1%/98.6% accuracy under five activities when trained with synthesized data only/adding a small amount of real data. For skeleton tracking, PmSyn achieves an overall skeleton tracking error of 10.5 cm/6.2 cm when trained with synthesized data only/adding a small amount of real data.