In this research, we investigate the application of the Audio Spectrogram Transformer (AST), a novel attention-based model, in conjunction with two advanced data augmentation techniques, ConnectMix and NEFTune, for enhancing respiratory sound classification. Our approach adapts the AST, originally designed for image classification, to analyze and classify respiratory sounds, leveraging its capability to capture long-range dependencies and intricate patterns in audio data. ConnectMix, a method that maintains the spatial continuity of audio spectrogram patches, and NEFTune, which introduces controlled noise during training, are integrated to overcome challenges associated with the ICBHI 2017 dataset’s variability and limited size. Our comprehensive evaluation demonstrates that our proposed model excels in binary and multi-class tasks for respiratory sounds, surpassing traditional methods in accuracy while also ensuring greater model reliability across diverse clinical scenarios. The enhanced model performance underscores the potential of these methodologies to revolutionize respiratory sound diagnostics, offering insights that could lead to advancements in automated medical analysis tools.