Semantic communication serves as a novel paradigm and attracts the broad interest of researchers. One critical aspect of it is the multi-user semantic communication theory, which can further promote its application to the practical network environment. While most existing works focused on the design of end-to-end single-user semantic transmission, a novel non-orthogonal multiple access (NOMA)-based multi-user semantic communication system named NOMASC is proposed in this paper. The proposed system can support semantic tranmission of multiple users with diverse modalities of source information. To avoid high demand for hardware, an asymmetric quantizer is employed at the end of the semantic encoder for discretizing the continuous full-resolution semantic feature. In addition, a neural network model is proposed for mapping the discrete feature into self-learned symbols and accomplishing intelligent multi-user detection (MUD) at the receiver. Simulation results demonstrate that the proposed system holds good performance in non-orthogonal transmission of multiple user signals and outperforms the other methods, especially at low-to-medium SNRs. Moreover, it has high robustness under various simulation settings and mismatched test scenarios.