Language identification (LID) systems are widely used in speech analytics, multilingual routing, and spoken-language interfaces, but their performance often drops in real-world conditions because of domain mismatch, including background noise, reverberation, channel effects, codecs, and far-field recording. We propose an embedding-level generative enhancement method based on flow matching that refines language embeddings from a pre-trained LID model. During training, we build paired embeddings from clean utterances and their distorted versions, and learn a continuous-time transformation that maps corrupted embeddings toward clean ones while preserving language-discriminative structure. At inference, any embedding is passed through the learned flow to obtain a refined, more robust representation. The method requires no changes to the LID pipeline and no language labels, while improving LID accuracy under challenging conditions without hurting standard performance.