Semantic Adaptation Transmission

• Semantic adaptation transmission is a design principle that selectively transmits task-relevant semantic information rather than full signal fidelity.
• It employs adaptive encoder and receiver mechanisms, including task-specific feature extraction, conditional rate control, and feedback-driven refinement.
• This approach achieves significant bandwidth savings and enhanced performance across modalities such as speech, image, video, and multimodal communications.

Semantic adaptation transmission denotes a class of semantic communication methods in which transmitted content, protection strength, computation, and reconstruction are adjusted according to the intended task, the semantic importance of different source components, and the prevailing communication conditions. In recent work, this idea appears as task-specific feature transmission, semantic entropy-based feature selection, adaptive computation, conditional rate control, feedback-driven refinement, and cross-layer protection, spanning speech, text, image, video, multimodal, and multi-hop settings (Han et al., 2022, Zhou et al., 2021, Gao et al., 2024). A common premise is that communication efficiency improves when systems transmit only information necessary for the downstream goal, rather than all source symbols or all perceptual details.

1. Conceptual basis

A recurrent formulation in the literature is that semantic communication should preserve meaning relative to a transmission goal, not necessarily full signal fidelity. In speech transmission, one system transmits only the semantic-relevant information for speech recognition and supplements it, when speech reconstruction is required, with a compact additional set of semantic-irrelevant but speech-related information such as duration, pitch, and power (Han et al., 2022). In unified multi-task transmission, domain adaptation is used so that only task-specific features, rather than all extracted features, need to be sent (Zhang et al., 2022). In masked auto-encoder-based image transmission, only the most semantically relevant patches are transmitted, and the number of retained patches is further constrained by channel capacity (Chen et al., 12 Feb 2025).

This task-relative view is formalized explicitly in the notion of semantic entropy. SemEntropy defines semantic entropy as the minimum expected number of semantic symbols needed to achieve the communication task goal:

$$H(X;Y) \triangleq \min_{E_S} \mathbb{E}\left(\text{dim}(Code^{E_S}(X))\right) \quad \text{s.t.}\quad P(Y|Code^{E_S}(X)) = P(Y|X).$$

This definition places the communication objective at the semantic level, not at the bit or symbol level (Rong et al., 2024).

A central implication is that “semantic” is not identical to “compressed.” Rather, semantic adaptation is selective: systems attempt to preserve the information that determines recognition, translation, classification, reconstruction, or decision-making, and discard information that is redundant for the current objective. The exact boundary between semantic-relevant and semantic-irrelevant information is therefore task-dependent, and in several systems it changes with channel conditions, task requests, or downstream feedback.

2. Encoder-side adaptation mechanisms

Many encoder-side designs implement semantic adaptation by estimating importance and allocating representational or channel resources accordingly. In deep speech semantic transmission, a learnable entropy model on the semantic latent space estimates importance diversity among semantic feature embeddings, and semantic features of different importance are allocated with different coding rates. The bandwidth used for each feature is determined by

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$

while a channel signal-to-noise ratio adaptation mechanism allows a single model to operate over various channel states (Xiao et al., 2022). In wireless deep video semantic transmission, entropy estimates are derived from hierarchical, spatial, and temporal priors, and the channel bandwidth for each latent embedding is adapted as

$k_{t,i} = \eta_t\, r_{t,i},$

so that resources are distributed within and among video frames according to temporal predictability and semantic novelty (Wang et al., 2022).

A second line of work adapts the structure of the encoder itself. In semantic-preserved speech communication, the transmitter uses a Semantic Encoder built from a VGG Module + BLSTM, followed by a Soft Alignment Module (Attention-based) and a Redundancy Removal Module. For speech-to-speech transmission, a CTC alignment module extracts semantic-irrelevant but reconstruction-relevant features such as duration, pitch, and power (Han et al., 2022). In text semantic communication, the Universal Transformer introduces an adaptive circulation mechanism with Adaptive Computation Time, so that each token can be processed for as many cycles as needed before halting. The transmitted symbols follow

$x = C_{\alpha}(S_{\beta}(s)),$

and the halting mechanism is governed by

$p_t = f_\varphi(x_t), \qquad h_t = h_{t-1} + p_t,$

with processing stopped when the cumulative halting score exceeds a threshold (Zhou et al., 2021).

Image-oriented systems increasingly combine semantic and channel awareness inside the feature extractor. An adaptive Mixture-of-Experts Swin Transformer replaces the standard feed-forward MLP with an Adaptive MoE MLP block, in which routed experts are activated dynamically according to both semantic content and real-time CSI. The output is written as

$$\mathbf{z}_t = \text{AD-MoE MLP}(\text{Norm}(\mathbf{z}'), \text{CSI}) + \mathbf{z}'.$$

This jointly driven routing breaks with single-driven designs that rely on content alone or channel state alone (Wan et al., 3 Apr 2026). Related masked auto-encoder systems use a multi-task-aware scoring mechanism and a channel-aware extractor to rank image patches and transmit only the highest-scoring patches that fit the current channel capacity, estimated by

$C = B \log_2(1 + \mathrm{SNR}).$

(Chen et al., 12 Feb 2025)

Across these designs, the transmitter is no longer a fixed compressor. It becomes a policy that decides what to encode, at what depth, with what rate, and sometimes with what expert subset, conditioned on semantics, task, and channel.

3. Receiver-side compensation, correction, and regeneration

Receiver-side adaptation is equally prominent. In semantic-preserved speech communication, the receiver uses a Channel Decoder, a Semantic Decoder + Beam Search, and a Semantic Corrector that employs an external pretrained LLM to enhance semantic correctness. For speech-to-speech transmission, a Speech Reconstructor combines predicted text with transmitted duration, pitch, and power using a FastSpeech2-based architecture and a pretrained GAN vocoder (HiFiGAN) (Han et al., 2022). This design separates semantic decoding from natural-speech regeneration.

Robust speech-to-text translation extends this principle further. Ross-S2T introduces a GAN-enabled deep semantic compensator that estimates lost semantic information within corrupted speech and a semantic probe-aided compensator that focuses repair on impaired positions. The clean-input path is

$$\mathbf{x} = {\mathfrak T}_C({\mathfrak T}_S(\mathbf{s})),$$

while the receiver produces

$$\widehat{\widetilde{\mathbf{t}}} = {\mathfrak T}_S^{-1}({\mathfrak T}_C^{-1}(\mathbf{y})).$$

The compensator is trained with adversarial, mean-squared, and probe-guided losses so that semantic fidelity and target-text understandability are improved under corrupted speech (Weng et al., 2024).

Generative semantic communication shifts adaptation toward iterative refinement. FAST-GSC parallelizes semantic extraction at the transmitter and inference at the receiver, reducing total latency from

$$\tau_{\text{total}} = \tau_{\text{ext}} + \tau_{\text{gen}}$$

to

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$0

Because parallelization can reduce task performance, the framework adds reinforcement learning-based temporal prompt engineering and a semantic difference calculation module with sequential conditional denoising to increase the impact of newly arrived semantic units (Wang et al., 2024). A related task-adaptive framework begins with a deep-compressed coarse semantic representation, lets the receiver generate task-specific feedback as a label or textual prompt, and then sends a masked edge update focused by CAM or CLIP-based attention. Reconstruction is performed by a Denoising Diffusion Probabilistic Model conditioned on both coarse and task-adapted semantics (Guo et al., 12 May 2025).

These receiver-side mechanisms suggest that semantic adaptation is often incomplete if it is confined to the encoder. In practice, many systems require post-channel semantic correction, semantic compensation, or conditional regeneration to recover task-relevant detail from aggressively compressed or partially corrupted semantic latents.

4. Cross-layer, protocol, and network-level adaptation

Several frameworks move semantic adaptation beyond the transceiver into transport, cross-layer control, and routing. The Cross-Layer Encrypted Semantic Communication framework for panoramic video transmission integrates feature extraction, encoding, encryption, cyclic redundancy check, retransmission, channel coding, and modulation. It assigns a semantic importance indicator to packets and adaptively selects the CRC polynomial, channel coding matrix, and retransmission limit:

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$1

(Gao et al., 2024)

Parallel-stream image transmission seeks compatibility with classical separate source-channel coding. ParaSC sends a conventional image stream and a lightweight semantic stream in parallel. The semantic branch is guided by the residual

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$2

and conditional rate adaptation is driven by a conditional entropy model on the semantic features. At the receiver, PAGNet dynamically aggregates the image and semantic streams according to SNR fluctuations (Xu et al., 30 Apr 2025). Emergency communication adopts a related hybridization: DA-ESemCom combines analog deep semantic coding with digital Distributed Source Coding, while a performance-constrained semantic coding model explicitly includes both semantic noise and channel noise and uses the Cramér-Rao lower bound as design guidance (Fu et al., 3 Jan 2025).

Protocol-level semantic adaptation appears in SPAT, which embeds source and destination ports directly into semantic representations rather than explicit headers. It uses differentiated semantic processing for uplink and downlink scenarios and an adaptive-rate controller that changes the number of transmitted semantic channels according to SNR and feature importance (Wang et al., 28 Apr 2026). At the physical layer, SemEntropy allocates more informative semantics to OFDM subcarriers with better channel conditions and uses semantic entropy to guide semantic key generation for secure transmission (Rong et al., 2024).

At the network layer, GraphJSCR models a Low Earth Orbit satellite network as a time-varying directed graph and jointly selects next hop, relay processing level, and semantic transmission budget. The relay adaptation is written as

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$3

where relay nodes may parse, prune, quantize, and repack semantic features en route (Zeng et al., 14 Apr 2026). This broadens semantic adaptation from source coding to end-to-end semantic flow control.

5. Multi-task, distributed, and multimodal extensions

A major development is the extension from single-task systems to unified multi-task and multimodal frameworks. U-DeepSC serves five tasks with multi-modalities in one model, using domain adaptation to separate shared and private features and a multi-exit decoder to allow simple tasks to exit early. The domain adaptation loss is

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$4

and the reported decoder depths are VQA (8), image retrieval (6), image reconstruction (4), text reconstruction (3), and sentiment analysis (2) (Zhang et al., 2022).

Distributed image transmission introduces another form of adaptation. The federated learning-based semantic communication framework equips each node with a hierarchical vision transformer-based extractor and a task-adaptive translator, while global aggregation improves semantic extraction and task performance. It also incorporates a CSI-based MIMO transmission module with SVD-based precoding and a two-step channel estimation procedure (Xie et al., 2023). In this setting, task adaptation is distributed across devices rather than concentrated in a single encoder.

Multimodal semantic communication increasingly abandons pixel-level reconstruction entirely. The Object-Attribute-Relation hierarchy constructs a decision-oriented topological graph from visual, textual, and audio streams and transmits object anchors, relations, and attributes according to semantic priority and bandwidth availability. The selection problem is formulated as

$$\bar{k}_{y_i} = Q\!\left(-\eta_y \log p_{y_i|\mathbf{z}}(y_i|\mathbf{z})\right),$$5

Objects are preserved first, then relations, then attributes, while text and audio priors compensate for severe visual degradation (Li et al., 9 Apr 2026).

These systems indicate that semantic adaptation transmission is no longer confined to source modality compression. It also encompasses task multiplexing, early exit, federated aggregation, hierarchical semantics, and cross-modal compensation.

6. Evaluation, trade-offs, and interpretive issues

Reported gains are substantial, but they are measured against different task objectives. In semantic-preserved speech transmission, the proposed method sends only 16% of the transmitted symbols required by the existing methods while achieving about 10% reduction in WER for speech-to-text transmission, and only 0.2% of the transmitted symbols required by the existing method for speech-to-speech transmission (Han et al., 2022). DSST saves up to 75% of channel bandwidth costs when achieving the same quality (Xiao et al., 2022). CLESC reduces bandwidth consumption by 85% compared to traditional cross-layer transmission schemes under equal visual quality conditions (Gao et al., 2024). FAST-GSC realizes a 52% reduction in residual task latency that extends beyond the fixed inference duration (Wang et al., 2024). The O-A-R hierarchy achieves over a 90% bandwidth saving and an 89% reduction in end-to-end latency while maintaining superior scene-graph accuracy under 1-3 kbps and eliminating the cliff effect in deep fading channels (Li et al., 9 Apr 2026). SemEntropy can keep the semantic accuracy remain 95% with 60% less transmission (Rong et al., 2024).

The metrics used to support these claims are correspondingly diverse: WER, sentence similarity, MCD, MOS, BLEU, STS, PESQ, MUSHRA, PSNR, SSIM, LPIPS, WS-PSNR, WS-SSIM, CLIP score, mAP@50, mAP@50:95, scene-graph recall, Graph Edit Distance, and latency all appear in the surveyed systems. This heterogeneity implies that performance comparisons across semantic communication papers are often task-bound rather than universal.

Several recurring misconceptions are therefore not sustained by the literature. First, semantic adaptation does not always mean abandoning perceptual reconstruction. Some systems optimize speech or image recovery quality directly, while others target classification, translation, retrieval, or scene-graph accuracy. Second, “semantic-only” transmission is not always sufficient: speech-to-speech reconstruction explicitly benefits from semantic-irrelevant but reconstruction-relevant information such as duration, pitch, and power (Han et al., 2022). Third, semantic communication is not necessarily incompatible with conventional stacks; CLESC, ParaSC, SPAT, and DA-ESemCom all retain compatibility with classical communication components or protocols (Gao et al., 2024, Xu et al., 30 Apr 2025, Wang et al., 28 Apr 2026, Fu et al., 3 Jan 2025).

A plausible implication is that semantic adaptation transmission is best viewed not as a single architecture but as a design principle. The principle is to match semantic representation, protection, and reconstruction effort to the downstream task and the current communication context. The surveyed work shows this principle appearing at multiple layers simultaneously: feature extraction, channel budget allocation, protocol design, retransmission control, routing, generative refinement, and multimodal fusion.