<aside> 🎵 Example Song Perturbations (Section 5.1 in Paper)
</aside>
Exploring musical roots: an audio walkthrough.
Subjective Listening Evaluation Examples
We assume that any generative music model will add some degree of variation to a training example during the generation process—the aim of these models are not to replicate the training data exactly. This variation could take many forms such as changing the pitch, speed, melody, etc.. Therefore, in this section we evaluate the ability of our methodology to return target songs that have been modified by given perturbations. For varying amounts of each perturbation, we evaluate how frequently the target song (the unmodified clip) is returned as the most similar, within the top 5 similar songs, and within the top 10 most similar songs. The 7 types of perturbations we evaluate are:
You can click any of these perturbations above to quickly jump to below examples.
We selected these because we envision them as common alterations to music that would not render it unrecognizable by a human listener. We are not seeking to evaluate all types of adversarial noise since we are assuming users and creators are working cooperatively with these generative models to create something novel---not acting maliciously.
A common perturbation to audio that involves raising or lowering the original pitch of an audio clip without adjusting the length of the clip.