Samples. Tracks. MIDI.One vault that gets it.
AudioVault indexes three libraries with one engine — detecting key, scale, BPM and energy with pure-Go DSP, tagging everything with the AI provider of your choice, and then building the thing you actually wanted: a drum rack, a DJ set, a MIDI pattern. Local SQLite index, no telemetry, every network integration opt-in.
14-day free trial · no credit card · one-time purchase · all future updates included
Thousands of Files, Zero Structure
Splice, Loopcloud, NI, MPC packs, promo downloads, MIDI bundles, field recordings — folders pile up and the right file is impossible to find
No Metadata
Filenames like bounce_03_final.wav tell you nothing about key, tempo, or character. You audition dozens of files to find one.
Out of Key, Out of Tempo
A loop that clashes with your track kills momentum, and a record that clashes with the one before it kills a set. Checking by ear, file by file, does not scale.
Duplicates and Fakes
The same kick lives in five packs under five names, the same record sits on disk twice, and half the “FLAC” folder is a transcoded MP3.
Samples, Tracks and MIDI — One Engine
Three fully independent indexes with their own tables, scans and similarity search, sharing the same DSP, the same tagger and the same browser
One-shots & loops
Recursive worker-pool scanning with content hashing for incremental re-scan, and a virtualized table that stays smooth across very large libraries.
- WAV, AIFF, FLAC, OGG, MP3 and MP4/M4A/AAC
- Key, scale, BPM, timbre, attack, loop vs one-shot
- Auto-detects Splice, Loopcloud, Ableton, Logic, NI, MPC
- Sample-accurate waveform trim, in place or as a copy
Full records, for DJs
A separate index for finished songs, with the analysis a DJ actually needs rather than the one a producer needs.
- 1–10 energy score from RMS, brightness, tempo and ZCR
- Structural cues: intro, drop, breakdown, outro — editable
- “Meaningful preview” offset, so previews skip the intro
- Embedded or looked-up artwork, per-track memos
Patterns & progressions
A third index over your .mid packs with real content analysis, not just filename parsing.
- Pitch-class key detection straight from the note data
- Drum / melodic / chords / mixed classification
- Polyphony and chord-progression summaries
- Click to play through a built-in synth with a piano roll
Each library gets its own Explore view — a PCA map of everything you own, a dashboard, a key wheel and taxonomy panels — so the shape of your collection is something you can actually look at.
Pure-Go DSP, Computed on Import
No external services and no Python sidecar — key, tempo, timbre and perceptual embeddings are computed in-process the moment a file lands
Tonality Detection
An HPCP chroma front-end correlated against Krumhansl-Schmuckler profiles across all 24 major/minor rotations, with a bass-note tonic tie-breaker. Returns key, scale and a confidence figure.
Tempo, Grid & Energy
Spectral-flux onset envelope → autocorrelation → comb scoring, with half/double-time correction and loop-grid snapping. Tracks additionally get an energy score and structural cue points.
Timbre & Envelope
Spectral centroid (brightness), rolloff, flatness, bandwidth, contrast and 13 MFCCs — plus attack time, RMS, zero-crossing rate and a loop / one-shot heuristic.
Perceptual Embeddings
48-dimensional vectors split into timbre, rhythm and pitch groups, powering weighted-cosine similarity search across the whole vault — with the weights exposed as sliders.
See Whether That FLAC Is Really Lossless
A Spek-style spectrogram rendered at the file's native sample rate over its full length — because a lossy encoder's lowpass shows up as a razor-sharp horizontal edge
- Read it by eye. Content up to Nyquist means lossless; a hard shelf around 16 kHz means it was 128 kbps once, whatever the file extension says.
- Or let it tell you. The cutoff is measured off the long-term average spectrum, never a max-hold — one click or edit point sprays broadband energy into a single window and erases the shelf, but the band a codec discarded is exactly zero in every window.
- Deliberately conservative. The bitrate table is only consulted for a genuinely sharp shelf, because plenty of real masters simply fade out by 18 kHz and labelling those “lossy” would flag half a library.
- Fake-lossless is harder still. That accusation needs a lossless container, a very steep drop, and at least half a minute of audio — a short one-shot is legitimately dark up top.
Categorized in Plain Language
Bring your own API key or use your Claude subscription CLI — AudioVault turns metadata plus detected features into a clean, navigable tag tree
- Any provider. Anthropic, OpenAI or Gemini API keys — or the authenticated
claudeCLI, billed against your own plan with no API key at all. - Two models, two budgets. Tagging runs on the cheap classification model; the agents that build kits and sets get an independent, reasoning-grade one. Making the agents smarter costs nothing per sample.
- A real taxonomy. Nine dimensions — category, genre, character, mood, effects, playing technique, song structure, gear, type — with aliases, driving both the prompt vocabulary and the colored tag tree.
- Heuristic fallback. No provider configured? A deterministic filename-plus-feature heuristic still tags everything, entirely offline.
- Natural-language selection. Type “select all dark 808s” and it resolves against your real library vocabulary, not a guess.
Find the Sound You Hear in Your Head
Similarity search, reference matching, harmonic compatibility, and a map of your entire collection
Sounds-Like Search
Weighted nearest-neighbour over the perceptual embeddings, tunable across timbre, rhythm and pitch. Pick anything, get its closest cousins instantly — samples or tracks.
Query by File
Drag in an external file — even one that isn't in your vault — and surface the closest matches you already own. It is analyzed without being imported.
Match Track
Analyze a reference and get samples that fit by compatible key (Camelot / circle-of-fifths), BPM window and category — ready to layer.
Explore Views
A 3D Sound Galaxy for samples and PCA maps for tracks and MIDI, each with a dashboard, key wheel, taxonomy panel and feature-insight charts.
Deep Filtering
Key and scale, BPM range, brightness, duration, energy, loop vs one-shot, favorites, tags, path segments and full-text search — combined, across any of the three libraries.
Duplicates & Same-Song
Group samples by content hash to reclaim disk space. For records, a stronger song-identity check catches the same song stored as two different files.
From Vault to Session in One Click
Turn a selection — or a sentence — into a playable instrument, a mixed set, or a pattern
DAW Kits That Are Actually Playable
Generate an Ableton Drum Rack (.adg) or Live Set (.als), or a Maschine group (.mxgrp) or project (.mxprj), on 8 to 64 pads with optional portable sample copying.
One Kick, Not Ten
The default layout reads each pad's role from the taxonomy tags, places one representative per role in canonical kit order, then fills the spare pads with the variant whose embedding is farthest from what that role already has — contrast instead of near-duplicates.
Named, Editable, Re-exportable
A kit is a real entity with export history. Leave it live and the pads re-resolve from its filter on every export; freeze the layout and it re-exports identically forever — even after a re-scan renumbers your library.
Magic Sort & Set Generation
Playlists with named chapters, a greedy harmonic / BPM / energy chain to arrange them, next-track suggestions, set completion, shuffle variations, and ordered USB export with numeric prefixes.
Nine Genres, Or Just Ask
Techno, psytrance, house, tech house, minimal, DnB, breakbeat, hip-hop and trap templates with swing, humanize, variation and seed — exported as .mid, or as an .als whose drum rack is already bound to your own samples.
Generate What You Don't Have
ElevenLabs text-to-speech takes and text-prompted sound effects, rendered straight into the library as mono WAVs — then analyzed and tagged like anything you scanned.
Describe It. Watch It Work. Then Edit It.
The kit and set generators are tool-using agents over your real library — and they show you every step
- Grounded in your library. The model can search samples or tracks, ask for similar ones, check key compatibility and read a library overview. It cannot invent files: every id in its plan is validated against rows that actually appeared in a tool result, and anything hallucinated is dropped.
- Always a usable result. Pads the model left empty are completed by the deterministic curator, and a set brief with no AI provider configured falls back to a heuristic spec plus Magic Sort. The feature never simply disappears.
- Reviewed, not applied. A generated set arrives as a draft — named chapters, ordered tracks, per-track notes — that you read before it becomes real. A generated kit lands in the same editable pad grid you would have filled by hand.
- No duplicate songs.Because the index is path-keyed, one record kept as two files is two rows. Set generation de-duplicates on song identity, not row id, so the same track can't appear twice.
Local by Default, Honest About the Rest
The index, the audio and the DSP never leave your Mac, and there is no telemetry of any kind. Everything that does touch the network is a feature you switched on
Single signed binary
One notarized macOS app with an embedded UI and a local SQLite index. No Docker, no sidecar, no account to create, nothing to install alongside it.
No telemetry
No analytics, no crash pings, no usage reporting. Your library lives under ~/Library/Application Support/AudioVault and stays there.
Your own AI account
Tagging uses your API key or your Claude subscription, so the billing relationship — and the data-retention policy — is between you and the provider.
Take Back Your Library
Try it free for 14 days. Buy once for €59, own it forever — all future updates included.