Rate-Adaptive Vector Quantization for Deep Joint Source-Channel Coding

Abstract

Traditional deep learning-based joint source-channel coding (DeepJSCC) frameworks typically require separate models for different code rates and channel conditions, and remain incompatible with practical digital communication systems due to their analog transmission nature. To address these limitations, we propose a unified Rate-Adaptive Quantization-based DeepJSCC (RAQJSCC) framework that enables channel-aware rate adaptation within a single model while maintaining compatibility with conventional digital communication systems. The proposed framework achieves rate adaptivity through a rate-adaptive quantization (RAQ) mechanism that enables flexible codebook adaptation within a single model, supporting variable-rate operation without retraining. To improve robustness under vector quantization-based digital transmission, we design a cosine similarity-based codebook reordering (CSCR) mechanism that aligns quantized representations with modulation schemes and reduces the impact of index errors. Furthermore, we incorporate an adaptive modulation and coding (AMC) unit that jointly selects modulation orders and effective codebook sizes based on channel conditions, enabling dynamic link adaptation across diverse SNR regimes. Experimental results demonstrate that the proposed RAQJSCC framework outperforms existing DeepJSCC and separation-based methods in terms of reconstruction quality over a wide range of SNRs, while effectively mitigating the cliff effect. These results highlight the potential of RAQJSCC as a practical and scalable solution for next-generation wireless communication systems.