[marssaxman]

There's no info yet partly because I've been trying a variety of approaches, and I'm not sure yet which approach will work out best. The core of the tool is a scanner which extracts an audio feature vector for each track in your library. Armed with this feature matrix, we can apply clustering algorithms - the most successful so far has been a Gaussian mixture model. I'm currently working on a system which will hopefully improve accuracy by bootstrapping a feature selection model, using metadata tags as an initial ground truth for music similarity, thereby allowing us to reduce the actual number of features which need to be compared.

I started out imagining that this tool would continuously update the contents of my crates, but now I think I want it to be more of a manual process. I'm imagining it as an analysis and reporting tool more than an organizer; I'll ask it to identify the outliers in a given crate, or ask for suggested additions, then choose how to arrange things myself. This way, I can use the manually-curated organization of my library as additional training data for the similarity model.