Your cues go where
the room sings.
No rekordbox involved.

The collection on the left, the tracks in the middle, the track you are looking at on the right, the waveform underneath and the drive along the bottom. There is no import mode, no export mode and no second application: preparing a record and putting it on a stick happen in the same place.

Cue points placed by energy can only ever find the loud part. That is worth having — the drop is where most mixes turn — but the moment you actually reach for is usually the line the crowd sings, and a hook and the verse before it are the same loudness, the same instruments and the same key. Nothing in a spectrum tells them apart. The words do.

So Booth renders the vocal stem, runs it through a speech recogniser, and reads what comes back for repetition. Lines that say the same thing are grouped loosely, because a recogniser writes one sung phrase four slightly different ways and matching word-for-word would count a hook sung eight times as eight different lines. The group with the most separate airings is the hook, and it becomes a cue where it first lands.

Two things it refuses to do. An ad-lib is not a hook — a line needs at least two words, because “yeah” is the most repeated thing in half the vocal stems ever recorded and marks nothing. And a recogniser stuck in a loop does not invent one: airings less than four seconds apart count once, which is true of a chorus that sings its line twice over and true of Whisper emitting “thanks for watching” forty times over a breakdown.

Where the words came from, and how sure anything is

Before anything expensive, Booth asks a lyrics server. A hit costs one request where listening costs a stem separation and a recogniser pass — and what comes back is what somebody wrote down rather than a machine's guess at a vocal. Where the track was named by its own sound, or where the lyric agrees with what was already heard, it is taken outright. Where it is plausible and not certain — a remix is a different length from the record it is built on and keeps only some of its words — it is put up for a yes or a no, with both sets of words on screen, and you can correct them by hand.

The times a lyric arrives with are not trusted as this pressing's: an LRC is synced against one release and a DJ's library is extended mixes, so the words go in needing placing and are moved onto singing measured off the stem. Each track then says where its words came from — heard off the stem, looked up, or corrected by hand — because a recogniser's mistakes read like lyrics, and when a cue reads oddly that is most of what there is to go on.

And it says what it will cost first

Reading the words off a track that has no vocal stem yet means separating one, which is the slow part of this program and the part nothing else waits on. Asking for cues on a playlist is therefore asking for a separation per track in it — so the batch is described and priced before it starts, in minutes measured from the last separations this machine did rather than guessed from the hardware.

A player holds eight hot cues and a busy track offers more than eight moments. So the buttons are shared rather than ranked against each other — four to the arrangement, four to the words, then whatever the other kind did not want — and everything else worth marking becomes a memory cue, which costs no button at all.

They are named the way a DJ writes on the markers themselves: Start, then each section numbered within its own kind — Build 1, Drop 1, Break 1, Drop 2 — and End on the last bar. Numbering within the kind is the point; six markers all reading “drop” say nothing that looking at the waveform does not.

A sung line gets a verse number and the line itself, every time it comes round: V1 hold me closer now. A player shows a memory cue's comment while the track is loaded, so the words are on the deck in front of you and the number says where in the record you are. Cues from the words are rounded down to the beat rather than to the nearest one — a sung line rarely starts on the beat, and a hook cue that clips its own first word is one nobody presses twice.

A drive a player opens has two databases on it that have to agree about what is on the stick: the legacy export.pdb every deck since 2009 reads, and the newer exportLibrary.db that every player sold since 2023 needs. Booth writes both from one collection, so they cannot disagree, and every generation's analysis files go on — .DAT, .EXT and .2EX.

Then it reads the stick back and checks it. You say the oldest player the drive has to work on, and the checks use that generation's rules: anything it could not open — a FLAC on a nexus deck, a 96 kHz file on an NXS2 — is re-encoded to a 320 kbps MP3 on the way onto the stick, while your library keeps its lossless copy.

The audio on a USB stick is replaceable. The cues, the grids, the playlist order and the history a player wrote back after a gig are not, and they are a few files under PIONEER.

So every time a drive is written — or a prepared one is plugged in while Booth is running — those files are copied beside your library. Databases, analysis, artwork, settings, the lot. The audio is hard-linked to the library's own copy rather than duplicated, so archiving a 64 GB stick costs a few megabytes and a directory entry per track. Each copy is a plain folder with a manifest saying what was found: getting one back onto a stick is a copy, with no tool in the middle.

A drive is stored once per state. Leaving it plugged in does nothing; writing to it and plugging it in again stores the new state beside the old one. What counts as a state is the drive's own files — deliberately not the .DS_Store breadcrumbs macOS sprinkles on a stick the moment anything looks at one, which would otherwise make a drive that is different every time it is read, and copied every time it is read.

The repair that cannot be done afterwards

A track your collection names but the disk has lost — the greyed-out row — is put back, off the drive, at the path the collection already gives it. That one happens whatever your settings say, because a row that cannot be played is broken rather than absent: the names are in the collection, the music is on the stick, and once the stick is gone neither half answers for the other.

Music on the stick your library has not got is copied in by default — the expensive option, on purpose. Everything else here is measured in disk, which can be bought.

And what you actually played comes back

A player writes a history to the stick it played from: every track it loaded, in order, one session a night. It is the only thing on a drive a collection cannot produce for itself, and it is the record of what was played rather than what was prepared. When a drive is copied, its history is read back and becomes playlists, named by the date the player gave them.

rekordbox is not disliked because it is bad at its job. It is disliked because it is load-bearing: the lock-in is not the application, it is the format on the stick. That format is public, documented and writable — so the honest comparison is feature by feature, including the features that go the other way.

The macOS app is built by the project's own packaging script and attached to each release — the library window and the booth-cli command-line tool in one bundle, universal for Apple silicon and Intel, so the two can never be different versions of each other.

git clone https://github.com/jwilkins/booth
cd booth
cargo run --release          # the library window

Cueing from the words needs a speech recogniser installed separately — Booth does not ship one and will not download one — and stems need a separation backend. Both are pointed at in Settings, and everything else works without them.