A Joint Modulation System with Variable Transmission Rate for Semantic Communications

By leveraging the advantages of deep learning, semantic communication (SemCom) has a superior spectral efficiency. However, the existing SemCom systems are limited by fixed-length encoding and analog transmission paradigms, and thus cannot fully demonstrate their efficiency advantages. This paper proposes a joint semantic modulation system based on the masked autoencoder structure (VJ-MAE). The system realizes variable transmission rates of semantic information by introducing a masking ratio control function based on source entropy and channel state estimation, along with an entropy-weighted mask unit sampling strategy. Meanwhile, a semantic modulation scheme is designed to ensure that finitely scalar-quantized semantic features share the same distribution as the constellation points, with layer normalization applied to constrain them into an approximately probabilistic-shaped distribution. Furthermore, encoding and modulation can be carried out through end-to-end joint optimization using gradient straight-through estimation. Experiments show that the reconstruction effect and robustness of VJ-MAE under different channel conditions are superior to the traditional separation coding and fixed-length coding schemes.