This demo provides an interactive exploration to understand how using different audio embeddings influence the FAD scores and what aspects of audio each embedding model captures. You can listen to the outliers—musical excerpts receiving the five highest and lowest per-song FAD for various combinations of audio embedding and reference sets. These insights are essential to discern the biases and characteristics captured by various embeddings in relation to musical genres and instruments, enabling a more nuanced understanding of how synthetic tones, distortions, and uniquely recognizable classes affect FAD scores. For a comprehensive exploration of these aspects and detailed anecdotal findings, you can learn more in our paper here: Adapting Frechet Audio Distance for Generative Music Evaluation.