.esca

3D sample library browser for streamlining discovery

problem

I have roughly 86,000 audio samples on disk that I use for producing music. Every tool built for a library that size assumes I already know what I am looking for, so the search box wants a filename, a tag, or a folder. None of those describe how a sample actually sounds. The result is a library I own and under-use. I reach for the same few hundred sounds and the other eighty-five thousand may as well not exist.

solution

.esca throws out the list. Every sample is placed in a three-dimensional space by its own sonic character, so proximity means similarity and the entire library becomes one continuous thing you can move through and hear in real-time. Finding a sound stops being recall and becomes navigation. Samples drag straight into the DAW, and where the library has a gap, a generative model fills it inside the same space rather than in a separate window. esca makes large audio sample sets glanceable.

A library you cannot hear is not a library.

The hard problem here is retrieval where the query is not a word. Text search works when you can name the thing. Nobody can name a sound. So the interface has to let you recognize rather than recall, which means the whole library has to be present at once and the distance between two points has to mean something real. esca solves this through visual organization of your entire sample library. Choose between UMAP or custom dimensions to re-plot samples and build collections of favorites with a simple mouse click.

Filter, search and hover - three steps to discover.

Beauty is a functional requirement, not a finish.

Sound design is hours of browsing. If those hours feel like file management, you stop early and the work suffers. The aesthetic, bioluminescence meets electronica, exists so that time inside the library is time you want to spend. That is a performance requirement dressed as a visual one.

Selected nodes are tethered to similar-sounding samples for serendipitous discovery

Generation belongs inside the space, not beside it.

Phase 1 of the project was to accurately plot samples as nodes that are glanceable. Phase 2 is generating like samples from selected "favorites". When the model makes a new sample it appears in the same coordinate system as everything else, surrounded by its neighbors. A generated sound arrives with context instead of alone in a blank panel, and it is judged the same way every other sample is judged, by where it sits and what it sits near.