Train Once, Deploy Broadly: Waveform Transformation for AWGN-Trained Semantic Communications

Bib

@ARTICLE{yoo2026trainonce,
  author={Yoo, Hanju and Ko, Minjae and Kim, Songkuk and Chae, Chan-Byoung and {Heath, Jr.}, Robert W.},
  journal={Submitted},
  title={Train Once, Deploy Broadly: Waveform Transformation for AWGN-Trained Semantic Communications},
  year={2026}
}

Hanju Yoo^†, Minjae Ko^†, Songkuk Kim^†, Chan-Byoung Chae^†, and Robert W. Heath, Jr.^‡

^† School of Integrated Technology, Yonsei University, Seoul 03722, South Korea
^‡ Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, CA 92093, USA

Scenario for train-once deploy-broadly semantic communications — The semantic model is trained once under AWGN, while waveform processing reshapes the practical OFDM deployment error before it reaches the decoder.

Abstract

Deep joint source-channel coding (DeepJSCC)-based semantic communication systems are typically trained under additive white Gaussian noise (AWGN), whereas practical wireless links produce error statistics that differ significantly from this assumption. This mismatch fundamentally limits their deployability, particularly in orthogonal frequency division multiplexing (OFDM) systems where the decoder-input error remaining after channel estimation and equalization is structured, correlated, and non-Gaussian. In this work, we investigate a decoder-aware framework that characterizes this mismatch via a second-order vulnerability model and motivates adapting the waveform, rather than the model, to the deployment environment. We propose a fixed waveform-level processing framework based on unitary spreading, phase dithering, and symbol permutation, which reshapes error statistics to better match the training assumptions. The proposed design reduces error correlation, promotes Gaussian-like behavior, and improves reconstruction quality by up to 0.9 dB in peak signal-to-noise ratio (PSNR) over raw OFDM. More broadly, it illustrates how waveform design can serve as a statistical interface between learned representations and physical channels, enabling a training-deployment decoupling paradigm for robust semantic communications without retraining or transmitter-side channel state information.

Why This Matters

The obvious fix for deployment mismatch is retraining, but that scales poorly. Every deployment condition, hardware chain, or propagation environment then becomes a model-management problem. This paper reframes the issue as an interface problem: the learned decoder expects AWGN-like perturbations, while practical OFDM delivers structured residual errors.

That viewpoint separates learning from deployment. If waveform processing can stabilize the error statistics seen by the decoder, the semantic codec can remain fixed while the physical interface absorbs part of the deployment variability.

What This Paper Does

The paper introduces a decoder-centric vulnerability model. Rather than measuring only average error power, it studies whether the deployment error covariance aligns with directions that strongly affect reconstruction loss. This explains why errors with the same ESNR can produce different semantic quality.

The waveform then targets three effects. Spreading averages diagonal error power and promotes Gaussian-like marginals. Phase dithering weakens coherent off-diagonal structure before mixing. Permutation disperses structured correlation away from decoder-sensitive local neighborhoods. The resulting design requires no retraining and no transmitter-side CSI.

PSNR versus ESNR under ZF equalization — Simulation with ZF equalization: waveform processing improves PSNR over raw OFDM at comparable effective SNR.

PSNR versus ESNR under MMSE equalization — Simulation with MMSE equalization: permutation-containing variants become more important when residual symbol-dependent structure dominates.

Prototype PSNR versus ESNR under ZF equalization — Over-the-air prototype with ZF equalization: the measured trend remains consistent with the simulation results.

up to 0.9 dBPSNR improvement over raw OFDM in practical links.

30.83 dBKodak ZF result for the DUP waveform near ESNR = 10 dB.

30.06 dBKodak MMSE result for UP near ESNR = 10 dB.

100 trialsOver-the-air prototype trials per waveform configuration.

Key Results

Practical OFDM errors decompose into colored equalized noise and symbol-dependent residual equalization terms.
Semantic distortion depends on alignment between deployment error covariance and decoder vulnerability, not only on average ESNR.
Spreading makes the measured decoder-input error marginal substantially more Gaussian-like.
Permutation becomes especially useful under MMSE equalization, where symbol-dependent residual structure matters more.
Simulation and USRP measurements both show that waveform processing recovers quality without changing the semantic model.

Error Covariance Reshaping

The covariance maps show what the waveform is doing beyond improving a scalar PSNR curve. In raw practical OFDM, the decoder-input error has structured covariance: the total error mixes a symbol-dependent residual term and a colored equalized-noise term. After the DUP waveform and inverse permutation, diagonal power is smoother and structured off-diagonal components are dispersed, reducing harmful alignment with the decoder vulnerability structure.

Raw total decoder-input error covariance — Raw OFDM: total decoder-input error covariance.

Raw residual equalization covariance — Raw OFDM: structured symbol-dependent residual term.

Raw equalized-noise covariance — Raw OFDM: colored equalized-noise term.

DUP total decoder-input error covariance — DUP: total covariance after waveform reshaping.

DUP residual equalization covariance — DUP: residual structure is dispersed after inverse permutation.

DUP equalized-noise covariance — DUP: equalized-noise covariance is also less locally concentrated.