An analysis of dimensionality reduction and music information retrieval techniques for the visual representation of large audio datasets.

Abstract

Navigation and search has long been a bottleneck activity when working with sound. Recent advancements in dimensionality reduction techniques have made it possible to replace the classic directory tree with a spacial map of audio files, resulting in a more intuitive representation. I provide a method of evaluating visualizations using methods from spatial statistics and find that a combination of PCA + STFT + UMAP methods gives the best scoring, although unexpected 2D embedding on a dataset of instrument recordings.