The birth prevalence of congenital heart disease (CHD) is 9/1000, and it is always challenging for paediatricians to diagnose. We propose a novel lightweight approach for CHD detection using a combined 2D convolutional neural network (CNN), bidirectional long short-term memory (Bi-LSTM), and attention mechanism. The objective is to classify paediatric phonocardiogram (PCG) recordings into five distinct classes: normal (N), functional (F), pathological (P), ventricular septal defect (VSD), and atrial septal defect (ASD). The work also investigates two-category classification into N and P and three-category classification into N, F, and P. The model achieved 95.6%, 93.2%, and 83.7% for two, three, and five categories. The model complexity and number of trainable parameters are much lower than the existing technologies, and the performance is comparable to the previous works. This is the first-ever deep learning-based work to classify paediatric heart sounds into more than four classes. The work was done on a dataset collected by the researchers. The robustness of the model was evaluated by testing against the publically available five-category adult PCG dataset. The model’s performance surpasses existing methods, highlighting its potential as a reliable tool for CHD and other cardiovascular disease detection.