Fully Learnable Multi-Rate Quantization for Digital Semantic Communication Systems

We propose ConcreteSC, a digital semantic communication framework that eliminates massive codebooks through temperature-controlled concrete distributions. Unlike vector quantization (VQ), it offers a fully differentiable solution to quantization, allowing end-to-end training even under channel noise. A simple masking mechanism further enables single-model, multi-rate transmission without retraining. Simulation results on ImageNet under Rayleigh and Rician fading demonstrate that ConcreteSC consistently surpasses VQ-based baselines in structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). Moreover, computational complexity scales linearly with bit length, avoiding the exponential complexity of codebooks. These advantages highlight ConcreteSC’s robustness, flexibility, and reduced overhead for semantic communication in next-generation wireless systems.