Channel-Aware Discrete Semantic Coding

• Channel-aware discrete semantic coding is a framework that represents task-relevant information using discrete symbols and adapts its design to channel conditions.
• It employs mechanisms such as semantic unequal error protection, distribution shaping, and channel-aware codebook geometry to enhance communication robustness.
• Empirical evidence indicates improved metrics like PSNR, LPIPS, and classification accuracy compared to traditional schemes, demonstrating its practical efficacy.

A channel-aware discrete semantic coding framework is a semantic communication architecture in which task-relevant information is represented by discrete symbols—such as video tokens, codebook indices, variable-length bitstrings, or foundation-model tokens—and the design of source coding, symbol usage, modulation, and protection is explicitly matched to channel conditions rather than optimized in isolation (Men et al., 2 Mar 2026, Zhang et al., 6 Aug 2025, Wang et al., 8 Oct 2025). In this literature, the same discrete representation is expected to support semantic fidelity, robustness under noisy links, and compatibility with digital communication infrastructures, while remaining useful to downstream models and tasks (Zhou et al., 11 Nov 2025, Bao et al., 10 Jun 2026).

1. Defining characteristics

The phrase combines three requirements. “Discrete” means that communication operates on finite alphabets: video tokens as codebook indices, VQ codewords, token IDs of foundation models, or binary codewords of bounded length, instead of raw pixels or unconstrained continuous latents (Men et al., 2 Mar 2026, Zhou et al., 11 Nov 2025). “Semantic” means that the framework does not treat all source components equally; it distinguishes task-relevant or user-intended content from less important content, and it optimizes the representation with respect to semantic fidelity, task accuracy, or topological structure rather than only per-pixel distortion (Men et al., 2 Mar 2026, Wu et al., 2 Mar 2026). “Channel-aware” means that source precision, codebook geometry, symbol usage, modulation, channel coding, or retransmission policy is conditioned on channel state, channel transition probabilities, SNR, or environment-derived channel knowledge (Zhang et al., 6 Aug 2025, Meng et al., 21 Oct 2025).

A notable feature of this research area is that “channel-aware” does not denote a single mechanism. In some systems it means explicit mixed-integer optimization of bit precision and modulation-and-coding scheme under BLER constraints (Men et al., 2 Mar 2026). In others it means shaping the empirical distribution of codeword activations toward a channel-designed optimal input distribution via Wasserstein regularization (Zhang et al., 6 Aug 2025), or introducing a channel-aware semantic distortion term into codebook training (Wang et al., 8 Oct 2025). STCC extends the idea to token-level constellations, where channel topology is aligned with the semantic embedding space so that channel noise produces “Semantic Drift” or “Structural Distortion” rather than random corruption (Bao et al., 10 Jun 2026).

This body of work also departs from a common misconception that semantic communication must be analog and end-to-end monolithic. Several representative systems are explicitly digital-first: they map codeword indices to K-QAM, operate on discrete memoryless channels, or optimize discrete modulation probabilities while remaining end-to-end trainable (Zhang et al., 6 Aug 2025, Bo et al., 2023, Meng et al., 21 Oct 2025). Video TokenCom, by contrast, keeps source and channel coding modular while still performing joint adaptation across semantic classes and PHY choices (Men et al., 2 Mar 2026).

2. Discrete semantic representations

The representational substrate varies, but the unifying principle is that the communication alphabet is discrete and semantically structured. In Video TokenCom, the video is converted into a 3D grid of discrete tokens by a pretrained tokenizer, producing integer indices on a spatio-temporal lattice; text-conditioned vision-language modeling and optical-flow propagation then partition those tokens into intended and non-intended classes (Men et al., 2 Mar 2026). Intended tokens are transmitted with full codebook precision, whereas non-intended tokens use reduced-precision differential coding. A representative rate expression is

$$B_{\text{tokens}} = thw\big(\rho_s B_{\text{full}} + (1-\rho_s)B_\Delta\big),$$

which makes the bit budget directly depend on the token-level intended ratio $\rho_s$ and the reduced differential precision $B_\Delta$ (Men et al., 2 Mar 2026).

In codebook-based digital semantic communication, the discrete unit is usually a learned semantic index. DeepJSCC-CDSC uses a learnable codebook $\mathcal{C}=\{c_1,\dots,c_K\}$, aligns codebook size with the modulation order, and maps encoder outputs to codeword selections that are then transmitted over K-QAM (Zhang et al., 6 Aug 2025). The theoretically-grounded codebook framework defines a codebook $$\boldsymbol{C}=[\boldsymbol{c}_1,\dots,\boldsymbol{c}_K]$$, uses nearest-neighbor quantization to generate a semantic index sequence $\boldsymbol{S}=[s_m]_{m=1}^M$, and interprets the Voronoi partition induced by $q(\boldsymbol{z}_m)$ as the engineering realization of a semantic equivalence-class partition (Wang et al., 8 Oct 2025). VQJSCC follows the same broad pattern on a discrete memoryless channel, but adds explicit channel transition probabilities to the codebook update rule (Meng et al., 21 Oct 2025).

A more radical discrete formulation appears in variable-length end-to-end coding. There the encoder is structurally decomposed into a length controller and a content encoder, and the transmitted semantic representation is a variable-length binary string with rate

$R \triangleq \mathbb{E}_Z[l(Z)].$

This makes bit-level communication rate a first-class optimization variable instead of an indirect proxy derived from latent dimension or bandwidth ratio (Zhou et al., 11 Nov 2025).

STCC pushes discreteness to the interface with foundation models. It treats WordPiece or visual tokenizer outputs as the source symbols, maps token embeddings through a learned source-channel semantic token codec into complex channel symbols, and recovers token IDs by nearest-neighbor search in the original embedding space (Bao et al., 10 Jun 2026). This makes the communication stack directly compatible with token-based LLMs and LVMs without receiver-side modification.

3. Mechanisms of channel awareness

The most explicit channel-aware design in the provided literature is unequal error protection over semantic classes. Video TokenCom groups tokens into intended and non-intended classes, fixes full precision for intended tokens, allows multiple precision candidates for non-intended tokens, and jointly selects per-class source precision, modulation order, and coding rate through a mixed-integer linear program (Men et al., 2 Mar 2026). The optimization minimizes a weighted sum of normalized distortion and delay under a global resource budget and per-class BLER constraints, with candidate configurations pruned by SNR. The result is semantic UEP: high-importance tokens receive stronger protection and larger alphabets, while low-importance tokens receive lower precision and weaker protection.

A second mechanism is distribution shaping. DeepJSCC-CDSC defines an empirical codeword activation distribution $\hat p(k)$ and a target distribution $q(k)$ derived from channel-capacity considerations, then penalizes a Wasserstein distance $\rho_s$0 in the training objective (Zhang et al., 6 Aug 2025). This regularizer simultaneously combats codebook collapse, increases effective alphabet usage, and aligns semantic symbol statistics with channel-optimal input distributions. In this formulation, channel awareness is not only about protecting symbols after quantization; it is already present in the marginal statistics of which symbols the semantic encoder chooses to activate.

A third mechanism is channel-aware codebook geometry. The theoretically-grounded codebook framework derives a channel-induced semantic distortion term

$\rho_s$1

combines it with quantization distortion and entropy regularization, and then studies the optimal codebook size through

$\rho_s$2

This makes the codebook itself sensitive to bit-flip probability, codeword distances, and index entropy (Wang et al., 8 Oct 2025). VQJSCC adopts a closely related but more operationally channel-specific view: its channel-aware vector quantization loss weights semantic distances by transition probabilities $\rho_s$3, so symbols that are easily confused by the channel are pulled toward semantically similar codewords (Meng et al., 21 Oct 2025). When codebook order and modulation order mismatch, it decomposes the stream into independently optimized subchannels and uses multiple codebooks aligned to those subchannels (Meng et al., 21 Oct 2025).

A fourth mechanism is learned probabilistic coding-modulation. JCM models the transmitter as a probabilistic encoder-modulator that learns transition probabilities from source data to discrete constellation symbols and trains them under an information-theoretic variational objective defined through the explicit AWGN channel law (Bo et al., 2023). Here the discrete semantic code is a sequence of constellation symbols, and channel awareness appears through the mutual-information objective and the resulting probabilistic shaping.

4. Representative systems and empirical evidence

The main instantiations differ in discrete unit, adaptation mechanism, and task objective, but they share the same architectural idea: a discrete semantic alphabet is co-designed with a noisy channel rather than merely protected after the fact.

Framework	Discrete semantic unit	Channel-aware mechanism
Video TokenCom (Men et al., 2 Mar 2026)	Video tokens / codebook indices	Semantic UEP with MILP over bit precision and MCS
DeepJSCC-CDSC (Zhang et al., 6 Aug 2025)	Learnable codebook indices	Wasserstein alignment to channel-designed input distribution
Theoretically-grounded codebook (Wang et al., 8 Oct 2025)	Quantized semantic indices	Entropy regularization and channel-aware semantic distortion loss
VQJSCC with CAVQ (Meng et al., 21 Oct 2025)	VQ codeword indices	Transition-probability-aware quantization and multi-codebook alignment
E2EC (Zhou et al., 11 Nov 2025)	Variable-length binary codewords	Policy-gradient optimization of semantic distortion plus bit length
STCC (Bao et al., 10 Jun 2026)	Foundation-model semantic tokens	Learned hyperspherical constellations with triple-loss topology alignment

Video TokenCom provides a particularly direct empirical demonstration. On UVG at BPP $\rho_s$4 for TokenCom versus $\rho_s$5 for VC-DM and H.265, the average PSNR is 26.36 for TokenCom, 24.47 for VC-DM, and 23.28 for H.265; LPIPS is 0.095 versus 0.104 and 0.184; FVD is 1289 versus 2087 and 4010 (Men et al., 2 Mar 2026). In channel-adaptive experiments, H.265 fails up to 4 dB, at 6 dB its PSNR is only 11.24 dB, whereas TokenCom maintains robust decoding through semantic UEP (Men et al., 2 Mar 2026).

DeepJSCC-CDSC targets inference rather than reconstruction. With a compact codebook $\rho_s$6, it achieves the highest accuracy at all SNRs; with $\rho_s$7, it maintains stable $\rho_s$8 accuracy across SNR for both $\rho_s$9 and $B_\Delta$0; and when increasing dimension from $B_\Delta$1 to $B_\Delta$2 with $B_\Delta$3, it reduces the negative impact of increased dimensionality by 63.46% versus DeepJSCC-RIB, 36.03% versus DeepJSCC-VIB, and 25.2% versus DeepJSCC-G (Zhang et al., 6 Aug 2025).

The theoretically-grounded codebook paper provides a clean ablation of channel-aware loss design. At SNR = 10 dB, its codebook achieves a 24.1% improvement in PSNR and a 46.5% improvement in LPIPS compared to the existing codebook designs (Wang et al., 8 Oct 2025). This is notable because the gain is attributed not to a larger backbone, but to entropy regularization and channel-aware semantic distortion during codebook training.

STCC shows what happens when the discrete symbols are foundation-model tokens. Under AWGN at 0 dB, STC_H achieves text token accuracy 86.91%, image token accuracy 82.39%, text semantic similarity $B_\Delta$4, and image LPIPS $B_\Delta$5; under fading at 0 dB, it yields 65.55% sentiment classification accuracy on SST-2, and for ImageNet classification it reaches 14.27% top-1 accuracy at 0 dB and 52.87% at 5 dB (Bao et al., 10 Jun 2026). The qualitative claim accompanying these numbers is central to the framework: channel errors are turned into semantic variations rather than catastrophic random substitutions (Bao et al., 10 Jun 2026).

5. Relation to adjacent paradigms

Not every channel-aware semantic communication method is an explicit discrete semantic coding framework, but several adjacent lines supply mechanisms that are directly reusable in one. TopoJSCC is continuous-valued and does not explicitly discretize semantics, yet its persistent-homology losses on image space and latent space are presented as directly relevant to discrete codebooks because they can preserve connectivity, loops, and latent manifold structure under channel perturbations (Erak et al., 17 Mar 2026). This suggests that a discrete codebook can be regularized not only for distortion and entropy, but also for topological invariants.

Adaptive-JSSCC is likewise continuous, but it contributes a semantic sampling ratio distribution map and an SNR-conditioned attention-based channel adaptive module. Its own discussion explicitly states that the architecture provides the channel-aware backbone and semantic importance mechanism needed for a channel-aware discrete semantic coding framework once explicit discrete representations are overlaid (Qi et al., 11 Feb 2025). The same is true of LCFSC over MIMO fading channels: it introduces non-invasive CSI fusion masking and a learnable mask ratio, then explicitly explains how those mechanisms can govern which discrete semantic symbols are produced or protected in a VQ-based design (Xie et al., 2024).

Environment-aware channel priors form another adjacent direction. The generative channel knowledge base learns a mapping from positions, global image features, and fine-grained semantic features to channel matrices, and injects the resulting channel knowledge into both encoder and decoder of a JSCC system (Long et al., 7 Apr 2026). The paper then argues that the same conditioning architecture can support channel-aware codebook design or symbol selection in a discrete semantic system (Long et al., 7 Apr 2026).

Goal-oriented unequal coding offers a task-centric counterpart. G-JSSCC does not use latent tokens or VQ codebooks, but it computes Shapley-value-based importance over image regions, allocates different source qualities and protection levels, and defines coding efficiency as $B_\Delta$6 (Wu et al., 2 Mar 2026). The region-level formulation translates naturally to discrete tokens, where “core,” “helpful,” and “negative” symbols could receive different quantization and protection policies (Wu et al., 2 Mar 2026).

6. Open problems and research directions

Several unresolved questions recur across the literature. One concerns the semantic representation itself. Video TokenCom relies on fixed pretrained tokenizers and a fixed CLIP-based selector, uses only two semantic classes, and employs an empirical distortion model $B_\Delta$7; the paper explicitly identifies richer semantic hierarchies, continuous importance scores, and joint training of tokenizer, selector, and allocator as open directions (Men et al., 2 Mar 2026). The theoretically-grounded codebook paper derives an optimal codebook-size criterion, but also notes that the full optimization over $B_\Delta$8 is not implemented in experiments (Wang et al., 8 Oct 2025).

A second unresolved issue is channel realism and scalability. Much of the discrete literature still uses AWGN, BSC abstractions, or modest fading models; larger configuration spaces, multi-user settings, MIMO, and stronger channel mismatch remain less explored than single-link settings (Meng et al., 21 Oct 2025, Zhou et al., 11 Nov 2025). STCC demonstrates strong robustness under AWGN and Rayleigh fading, but it also identifies channel-model mismatch, vocabulary scaling, and the absence of explicit instantaneous CSI at the transmitter as limitations (Bao et al., 10 Jun 2026).

A third issue is how to combine discreteness with richer semantic structure. TopoJSCC explicitly proposes applying persistent-homology regularization to codeword embeddings or VQ-based codebooks, while task-oriented frameworks suggest combining semantic distortion with downstream metrics such as classification or segmentation accuracy (Erak et al., 17 Mar 2026, Wu et al., 2 Mar 2026). A plausible implication is that future frameworks will not treat codebooks merely as compression dictionaries; they will also be shaped as structured semantic manifolds whose neighborhoods remain meaningful after channel perturbation.

Finally, there is an emerging question of temporal adaptation. Variable-length E2EC shows that semantic communication can optimize the actual expected number of transmitted bits through a stochastic length controller (Zhou et al., 11 Nov 2025). Cache-enabled generative JSCC shows that a dynamic semantic codebook can evolve across transmissions, achieving an average BCR of 1/224 and as low as 1/1024 for a single image by reusing previously transmitted semantic components (Tang et al., 18 Mar 2026). Together, these results indicate that a channel-aware discrete semantic coding framework need not be static: its alphabet, length, and redundancy can evolve with both channel conditions and accumulated semantic experience.