Algorithm 2 Stage II: Prior-Guided Affective Diffusion Optimization

Input: Frozen Modules: JSCC Encoder/Decoder ${\theta }_{jscc}$, Semantic Encoder ${\mathcal{E}}_c$;

Trainable Modules: Structural Prior Network ${\mathcal{P}}_{prior}$ (including Duration Predictor),

Diffusion U-Net ${ϵ}_{\theta }$;

Data: Speech dataset $\mathcal{D}$, Ground-truth Mel-spectrogram ${\mathcal{X}}_0$;

Hyperparameters: Loss weights ${\lambda }_a,{\lambda }_s$, noise schedule ${\beta }_t$.

Output: Optimized Generative Parameters ${\theta }_{gen}=\{{\mathcal{P}}_{prior},{ϵ}_{\theta }\}$.

1:Initialize ${\theta }_{gen}$ randomly.  2:repeat  3:   for each batch $(W,{\mathcal{X}}_0)\sim \mathcal{D}$ do  4:     Digital Stream: Extract & Quantize semantic features:  5:        ${\widehat{Z}}_c\leftarrow \mathrm{Quantize}\left({\mathcal{E}}_c\left(W\right)\right)$  6:     Analog Stream: Transmit affective features via frozen JSCC:  7:        $Z_a=\mathrm{Agg}\left({\mathcal{E}}_a\left(W\right)\right)$  8:        ${\widehat{Z}}_a={\mathcal{D}}_{jscc}\left(\mathrm{Channel}\left({\mathcal{E}}_{jscc}\left(Z_a\right)\right)\right)$  9:     Predict unaligned distributions: $\tilde{\mu }={\mathcal{P}}_{prior}\left({\widehat{Z}}_c\right)$ 10:     Monotonic Alignment Search (MAS): 11:        $A^{*}=\mathrm{MAS}(\tilde{\mu },{\mathcal{X}}_0)$ 12:     Get aligned semantic skeleton & durations: 13:        $\mu =\mathrm{Align}(\tilde{\mu },A^{*}),d_{mas}=\mathrm{Sum}\left(A^{*}\right)$ 14:     Compute Structural Losses: 15:        ${\mathcal{L}}_s=\parallel \mu -{\mathcal{X}}_0{\parallel }^2,{\mathcal{L}}_t={\parallel log{d}_{pred}-log{d}_{mas}\parallel }^2$ 16:     Sample time $t\sim [0,T]$ and noise $\xi \sim \mathcal{N}(0,I)$ 17:     Sample noisy state ${\mathcal{X}}_t$ via OU forward process anchored at $\mu$: 18:        ${\mathcal{X}}_t={\rho }_t{\mathcal{X}}_0+(1-{\rho }_t)\mu +{\delta }_t\xi$ 19:     Predict noise residual with Affective FiLM conditioning: 20:        $\widehat{\xi }={ϵ}_{\theta }({\mathcal{X}}_t,\mu ,t,{\widehat{Z}}_a)$ 21:     Compute Diffusion Loss: 22:        ${\mathcal{L}}_d={\parallel \widehat{\xi }-\xi \parallel }^2$ 23:     Aggregate Total Loss: 24:        ${\mathcal{L}}_{total}={\mathcal{L}}_d+{\lambda }_a{\mathcal{L}}_a+{\lambda }_s({\mathcal{L}}_s+{\mathcal{L}}_t)$ 25:     Update gradients: ${\theta }_{gen}\leftarrow {\theta }_{gen}-\eta {\nabla }_{\theta }{\mathcal{L}}_{total}$ 26:   end for 27:until convergence 28:return ${\theta }_{gen}$