GROOT fusion model detects fake heart sounds with new benchmark dataset

A new paper from researchers (Girish, Orchid Chetia Phukan, Mohd Mujtaba Akhtar, et al.) tackles the emerging threat of synthetic heart sounds generated by neural audio codecs (NACs). The team introduces SHAC (Synthetic Heart Sound Detection), a task focused on identifying fake phonocardiograms (PCGs). To enable this research, they release CARDIOFAKE, the first benchmark dataset comprising both real PCGs and those synthesized by state-of-the-art neural audio codecs. They benchmark spectral representations (MFCC, LFCC) and self-supervised learning (SSL) representations (e.g., WavLM) for detection, providing a rigorous baseline for the field.

The paper's key contribution is GROOT, a fusion framework that combines spectral and SSL features to exploit their complementary strengths. Experiments show that GROOT, specifically using MFCC and WavLM, achieves state-of-the-art performance on CARDIOFAKE, significantly outperforming individual representations and all competitive baselines. This work has been accepted to INTERSPEECH 2026, highlighting its relevance to the speech and audio community. The ability to reliably detect synthetic heart sounds is critical for telemedicine and remote diagnostics, where fake recordings could compromise patient safety or insurance claims.