Algorithm 1 Stage I: Training Procedure for the Analog Affective Link

Input: Frozen Modules: Emotion Encoder ${\mathcal{E}}_e$, Speaker Encoder ${\mathcal{E}}_s$;

Trainable Modules: JSCC Encoder ${\mathcal{E}}_{jscc}$, JSCC Decoder ${\mathcal{D}}_{jscc}$ (collectively ${\theta }_{jscc}$);

Hyperparameters: Learning rate $\eta$, batch size B, transmit power P.

Output: Optimized Parameters ${\theta }_{jscc}$.

1:Initialize trainable parameters ${\theta }_{jscc}$ randomly.  2:repeat  3:   for each batch $W\sim \mathcal{D}$ (batch size B) do  4:        Extract emotion vectors: $Z_e={\mathcal{E}}_e\left(W\right)$  5:        Extract speaker vectors: $Z_s={\mathcal{E}}_s\left(W\right)$  6:        Concatenate latent features: $Z_a=[Z_e\oplus Z_s]$  7:        Map to channel symbols: ${\tilde{S}}_a={\mathcal{E}}_{jscc}\left(Z_a\right)$  8:        Apply Power Normalization:  9:        $S_a=\sqrt{P}·{\displaystyle \frac{{\tilde{S}}_a}{\sqrt{\mathbb{E}\parallel {\tilde{S}}_a{\parallel }^2+ϵ}}}$ 10:        Sample channel state $h_a\sim \mathcal{CN}(0,1)$ and noise $n_a\sim \mathcal{CN}(0,{\sigma }^2)$ 11:        Transmission: $Y_a=h_a·S_a+n_a$ 12:        Neural Inversion: ${\widehat{Z}}_a={\mathcal{D}}_{jscc}\left(Y_a\right)$ 13:        Compute Reconstruction Loss: 14:        ${\mathcal{L}}_a={\displaystyle \frac{1}{B}}{\sum }_{i=1}^B{∥Z_a^{\left(i\right)}-{\widehat{Z}}_a^{\left(i\right)}∥}_2^2$ 15:        Update gradients: ${\theta }_{jscc}\leftarrow {\theta }_{jscc}-\eta {\nabla }_{\theta }{\mathcal{L}}_a$ 16:   end for 17:until convergence or max epochs reached 18:return ${\theta }_{jscc}$