Automated medical transcription systems have made eminent strides in recent years, yet challenges remain in faultlessly transcribing medical terminology, varied accents, and specialized language with inadequate labelled data. Traditional transcription systems depend predominantly on enormous, domain-specific datasets, that are often high priced and too time-consuming to acquire. Few-shot learning (FSL) is a paradigm that allows the models to generalize from a smaller number of examples, offering a very optimistic solution to this problem. This paper explores the implementation of few-shot learning processes to improve the performance of automated medical transcription systems, particularly in scenarios with scant data availability for training. We suggest a framework that combines pre-trained language models, meta-learning, and transfer learning to refine transcription accuracy with nominal labelled examples. Results conclude that few-shot approach outperforms traditional models, achieving a reduction in word error rate and better grasping many intricate medical terms. We achieved an accuracy of 92.50% using few-shot learning approach on medical transcriptions dataset. Our outcomes suggest that few-shot learning has the capability to revolutionize transcriptions of medical domain, offering a prestigious and scalable solution for applications in health care.