Intent detection is a critical task in building spoken language understanding (SLU) systems. We propose a novel semi-supervised Dual Real Input Generative Adversarial Network (DRI-GAN) with triplet loss to enhance the performance of this task. This method effectively leverages both labeled and unlabeled data to achieve superior representation learning. We extract and fuse text embeddings from three locally deployed pre-trained Large Language Models (LLMs) and adapt these embeddings for training the DRI-GAN with triplet loss. Our experiments demonstrate three key findings: (i) In noisy SLU environments, the proposed method outperforms the state-of-the-art by +1.44%. (ii) In zero-shot cross-lingual scenarios, our approach yields substantial accuracy improvements, achieving an absolute gain of 4.19% on MultiATIS++ dataset and 14.60% on MASSIVE dataset. (iii) it achieves higher accuracy without fine-tuning, significantly reducing computational load.