sDAC -- Semantic Digital Analog Converter for Semantic Communications

TL;DR

This work tackles the challenge of reconciling semantic communications with conventional digital modulation by introducing sDAC, a generalized bi-directional converter that translates continuous semantic encoder outputs into discrete bits and back without altering existing systems. It proposes a discrete training framework to enable differentiable optimization through non-differentiable quantization and binarization steps, and introduces the ASE metric to capture semantic-level distortion due to quantization, channel noise, and end-to-end processing. The contributions include the sDAC module, a discrete training strategy with a learnable codebook and quantization adapter, the ASE metric, and extensive experiments showing robustness across semantic models, tasks, modulation methods, and channel conditions. The results indicate that sDAC achieves strong interoperability with digital modulation while preserving semantic task performance, particularly in low-SNR regimes and task-oriented settings, suggesting meaningful practical impact for next-generation 6G-like systems.

Abstract

In this paper, we propose a novel semantic digital analog converter (sDAC) for the compatibility of semantic communications and digital communications. Most of the current semantic communication systems are based on the analog modulations, ignoring their incorporation with digital communication systems, which are more common in practice. In fact, quantization methods in traditional communication systems are not appropriate for use in the era of semantic communication as these methods do not consider the semantic information inside symbols. In this case, any bit flip caused by channel noise can lead to a great performance drop. To address this challenge, sDAC is proposed. It is a simple yet efficient and generative module used to realize digital and analog bi-directional conversion. On the transmitter side, continuous values from the encoder are converted to binary bits and then can be modulated by any existing methods. After transmitting through the noisy channel, these bits get demodulated by paired methods and converted back to continuous values for further semantic decoding. The whole progress does not depend on any specific semantic model, modulation methods, or channel conditions. In the experiment section, the performance of sDAC is tested across different semantic models, semantic tasks, modulation methods, channel conditions and quantization orders. Test results show that the proposed sDAC has great generative properties and channel robustness.

Figures (9)

• Figure 1: An abstract common semantic communication framework. Source data of various modalities get encoded by various semantic encoders and output continuous-valued data. These data are added with analog channel noise to simulate wireless noisy channels. At the receiver, these data get decoded by various semantic decoders for various semantic destinations, including traditional reconstruction and rising semantic tasks.
• Figure 2: The illustration of the sDAC-based semantic communication system. It is composed of original semantic modules, DAC modules and binary symmetric channels. The output of the semantic encoder $s$ gets quantized by an analog-to-digital converter and then transmitted through a BSC to simulate the effect caused by intermediate processes, including modulation, equalization, etc. At the receiver, the process above gets inverted by a digital-to-analog converter to output $\hat{s}$. The process above does not modify any module in the original semantic communication system or digital traditional communication system, ensuring it is a great generalization.
• Figure 3: The internal structure of sDAC. Analog-to-digital converter transforms continuous values into binary bits through the nearest neighbours calculation between the learnable codebook and the output of the quantization adapter, which is a group convolution to map one single data to $q$ symbols. The digital-to-analog converter realizes the inverse process of combining binary bits back to continuous values for further semantic tasks. The end-to-end communication process is powered by the abstract BSC to simulate the digital communication process.
• Figure 4: An illustration of different SNR effects under the same BER with binary quantization. It proves that it is not BER but semantic distance, also known as semantic error, that really counts. It also makes it hard to reflect the end-to-end communication performance with a single metric BER.
• Figure 5: The illustration of a basic analog semantic communication system. It is implemented to alleviate the influence caused by the semantic codec. Up and down arrows mean 2X up and downsampling. $K$ indicates the kernel size of convolution. This system is based on image source data and for image transmission tasks.