Explore your samples by sound, then build beats — all in the browser, no install.

Glean analyzes the actual audio of every .wav in a folder, lays them out in a 2D similarity map (sounds that sound alike sit together, colored by type), lets you drag across the map to audition by ear, and drop the ones you like onto an 8-track step sequencer with per-track filter, EQ, pan and pitch. Inspired by XLN Audio XO and Waves Cosmos (not affiliated with or endorsed by either).

🔗 Live demo: https://yonie.github.io/glean/ — zero install; open it and point it at a folder of samples. Everything runs locally; nothing is uploaded.

• Sound-based map — each sample becomes a ~17-dimensional timbre vector (envelope, spectral centroid/spread/rolloff/flatness, log-band energies, zero-crossing rate) and the whole library is projected to 2D with PCA. Similar sounds cluster; dot color is the sample's audio-classified type (kick, snare, hat, …) — not its filename.
• Drag-to-scan — hold the mouse and sweep across the map; it retriggers the nearest sound as you move, so you browse by ear. The layout is computed once over the whole library and stays put — folder/search just filter which dots show.
• Folder browser — a real folder tree with breadcrumbs; works on any folder structure.
• Clickable legend — toggle sound-types to filter the map by color.
• 8-track step sequencer — 16 steps, adjustable BPM, tap tempo, and a permanent row of rotary knobs per track (VOL · PAN · FILT · RES · LOW · HIGH · PIT). Each lane is colored by its sample's sound. Live Web Audio playback.
• Load / swap — click a track to browse the map and drop a new sound onto it.

No build step. Either use the live demo, or run locally:

# clone, then serve the app folder (any static server works)
cd app && python -m http.server 8000
# open http://localhost:8000  →  "Open folder…"  →  pick a folder of .wav files

Opening app/index.html straight off disk works in most browsers too; a static server just avoids the odd file:// script-loading restriction.

The commercial tools use larger learned feature sets / ML embeddings and a non-linear projection (t-SNE/UMAP-style). Glean uses a transparent, dependency-free pipeline: hand-built timbre features → linear PCA to 2D. It produces a real similarity layout (kicks, hats, snares, pads separate out) and is easy to read and hack, but won't be as surgical as a trained model. The dot's discrete color label is a heuristic (~50–60% on ambiguous one-shots — a clap can read like a snare); trust the position more than the label. Swapping in UMAP/t-SNE + MFCCs is an obvious next step.

app/                 the web app (no build step)
  index.html
  css/styles.css
  js/util.js         shared state + helpers + colors
  js/dsp.js          FFT, timbre features, PCA  (pure, reusable)
  js/audio.js        decode/audition + per-track FX chain
  js/app.js          UI: folder browser, map, sequencer, transport
tools/               Python scripts to PREP/sort a sample library (see tools/README.md)
scripts/shoot.mjs    Playwright harness that regenerates the screenshots
docs/                screenshots + build notes

• Pattern save/load and export (MIDI / audio).
• Optional t-SNE/UMAP projection and MFCC features.
• Per-step parameter locks; more than 8 tracks.
• File System Access API to re-read a folder without re-picking.

If Glean is useful to you:

MIT — see LICENSE.