Dialect identification (DID) is a challenging task due to the narrow feature space shared by related dialects, making it more complex than traditional language identification. This study explores the application of Vision Transformers (ViTs) for dialect identification by leveraging spectrograms as input representations. Pre-trained ViT models, including base and large variants, were fine-tuned on spectrograms derived from audio samples of four Khasi dialects and four British Isles English accents. The models were evaluated achieving test accuracies ranging from 95% to 98%. The results demonstrate that ViTs, particularly the large variant pre-trained on ImageNet-21k, outperform smaller models, achieving the highest accuracy of 98% for both datasets. This study highlights the effectiveness of transfer learning in adapting image-based ViT models for audio-based dialect identification tasks, offering a robust approach for applications in automatic speech recognition and language modeling.