This paper presents the main features of our submission to the TidyLang 2026 Challenge. Since the challenge involves speakers producing speech in two or more languages, it is crucial to disentangle speaker-specific and language-specific characteristics. To address this issue, we propose a novel speaker-aware Mixture of Experts (MoE) architecture for spoken language verification. Our system extracts high-fidelity audio representations using a fine-tuned Wav2Vec 2.0 model combined with an attentive pooling mechanism. The core innovation lies in a gating network that estimates the probability that two utterances are produced by the same speaker. This probability is then used to dynamically weight the outputs of two language experts based on Neural PLDA: one specialized in same-speaker trials and the other in cross-speaker trials. This architecture effectively decouples language variability from speaker-related noise, thereby enhancing the overall robustness of the verification system.