A Latent Diffusion Model for Heart Sound Synthesis

There are already many analyses of heart sounds used to develop diagnostic systems for the heart condition. However, the existing heart sound database state is not sufficient to train a large number of deep learning models and is not balanced – the normal heart sounds are always more present than abnormal heart sounds. Hence, we are interested in algorithms that can generate heart sounds as they can enhance the current database state. We enter the field of large-scale modelling in medical synthesis by proposing a latent diffusion model for heart sound generation, which can generate highly realistic heart sounds. We further guide the synthesis process through text prompts and labels, revealing the research field of prompted heart sound synthesis. In terms of experimental results, our proposed method achieves good results compared to existing mathematical models. At the same time, when the generated audio is used as an unlabelled dataset in semi-supervised learning, it shows an improvement effect on the classification model. This indicates that the heart sounds generated by the diffusion model bear great value.