breakdig

Find the drum breaks, acapellas and other bare sections in your own music, cut to the bar.

breakdig runs every track in a folder through Demucs source separation and a downbeat tracker, then keeps a small index of how loud the drums, bass, vocals and everything else are in each bar. After that you can ask for "drums only, at least 2 bars, 85 to 100 BPM" and get back exact bar ranges you can audition and export from the original files. Or let breakdig drill turn a setlist into a tempo-ladder practice file per song: a click count-in, the song at 70% speed, then 75% on up to 100%.

Everything runs locally. Nothing is uploaded anywhere.

Install

You need Python 3.11, ffmpeg on PATH, and for reasonable speed an NVIDIA GPU. On Windows:

winget install Python.Python.3.11
winget install Gyan.FFmpeg

On Windows the simplest route is the zip. Download breakdig-0.2.0-windows.zip from the GitHub release, unzip it somewhere permanent, and run install.bat. It makes a virtual environment next to itself, installs breakdig into it, and swaps in the CUDA build of PyTorch if nvidia-smi is present. It prints the full path to breakdig.bat to use afterwards; add the folder to PATH to type just breakdig.

Or with pip:

py -3.11 -m venv .venv
.venv\Scripts\activate
pip install breakdig
pip install --force-reinstall --no-deps torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

The last line matters. On Windows, pip resolves torch to the CPU-only build, and separation on the CPU is 10 to 30 times slower. The cu130 wheels need NVIDIA driver 580 or newer; with an older driver use cu126 in the URL instead. breakdig prints a warning when it is running without CUDA.

The first breakdig index downloads two models: htdemucs_ft (322 MB) into %LOCALAPPDATA%\breakdig\models, and the beat_this checkpoint (81 MB) into the PyTorch hub cache at %USERPROFILE%\.cache\torch\hub\checkpoints. If the Demucs download is cut short, the run stops and deletes the partial file, and the next run downloads it again.

Quick start

Index a folder. This is the slow part and it only happens once per track:

> breakdig index D:\Music\Breaks
Scanning 14 files...
0 already indexed, 14 to do. Index: C:\Users\you\AppData\Local\breakdig
Done in 189s: 14 ok.
Separated tracks took 13.5s each on average.

Run it again and it only looks at what is new. Ctrl+C is safe at any point; the next run carries on from the track it was working on. A second index run on the same index while one is going stops straight away and says so.

> breakdig index D:\Music\Breaks
Scanning 14 files...
14 already indexed, 0 to do. Index: C:\Users\you\AppData\Local\breakdig
Done in 0s: nothing new.

Find sections. With no options, find looks for drums alone:

> breakdig find
ID          ARTIST - TITLE                               BARS     START       END    BPM  CLEAN
-----------------------------------------------------------------------------------------------
13:107-108  Bebop - Fidget Kaos                             2  3:15.724  3:19.403  130.2  -34.0
14:165-166  Innyu - For The Hell Of It                      2  5:07.525  5:11.300  127.7  -30.4

2 section(s). CLEAN is the loudest unwanted stem minus the target, in dB; lower is cleaner.

--silence-db 25 loosens the silence rule (see below) and finds a little more:

> breakdig find --silence-db 25
ID          ARTIST - TITLE                               BARS     START       END    BPM  CLEAN
-----------------------------------------------------------------------------------------------
13:107-109  Bebop - Fidget Kaos                             3  3:15.724  3:21.264  130.0  -29.8
14:164-166  Innyu - For The Hell Of It                      3  5:05.649  5:11.300  127.7  -26.5
14:7-8      Innyu - For The Hell Of It                      2  0:11.273  0:15.062  127.8  -26.4

3 section(s). CLEAN is the loudest unwanted stem minus the target, in dB; lower is cleaner.

> breakdig find --no drums --limit 5
ID          ARTIST - TITLE                               BARS     START       END    BPM  CLEAN
-----------------------------------------------------------------------------------------------
7:25-28     Igor Leontyev - Remote District                 4  0:16.358  0:21.674  183.1  -55.7
10:1-16     Zipp - Coffee Break                            16  0:01.526  0:33.507   60.3  -55.5
2:56-57     Dog On Springs - Footloose (feat. Paul Whit~    2  1:31.346  1:34.819  137.3  -55.3
7:4-6       Igor Leontyev - Remote District                 3  0:03.016  0:06.369  182.8  -54.9
2:43-45     Dog On Springs - Footloose (feat. Paul Whit~    3  1:09.037  1:14.137  142.3  -52.3

5 section(s). CLEAN is the loudest unwanted stem minus the target, in dB; lower is cleaner.

Export everything a find would list by giving the same filters to export, or pick sections by ID with --pick 13:107-109 14:7-8 (the other filters are ignored then):

> breakdig export --silence-db 25 --out D:\Samples\breaks
wrote   D:\Samples\breaks\Bebop - Fidget Kaos - 3 bars - 130 BPM - 03.16.wav
wrote   D:\Samples\breaks\Innyu - For The Hell Of It - 3 bars - 128 BPM - 05.06.wav
wrote   D:\Samples\breaks\Innyu - For The Hell Of It - 2 bars - 128 BPM - 00.11.wav

3 file(s) in D:\Samples\breaks

Or do all of it in the browser:

> breakdig ui
breakdig UI on http://127.0.0.1:8765  (Ctrl+C to stop)

Drills

breakdig drill turns a setlist into one practice MP3 per song: a click count-in, the whole song at 70% speed, a one-bar gap, then 75%, 80%, 85%, 90%, 95% and 100%. Copy the files to your phone and practice with a tempo ladder the way Anytune's Step-It-Up trainer plays, but as plain files that any player runs, offline.

The setlist is a folder, loose files, or an .m3u/.txt list with one path per line (# lines are comments). Every input is indexed first with the same resumable index as breakdig index, so the tempo and the bar grid come from the analysis:

> breakdig drill D:\Gigs\Nov14 --out D:\Drills
Scanning 14 files...
14 already indexed, 0 to do. Index: C:\Users\you\AppData\Local\breakdig
Done in 0s: nothing new.

Zipp - Pour Quoi Royale  140.2 BPM
   70%    98.1 BPM  4:10
   75%   105.1 BPM  3:53
   80%   112.1 BPM  3:39
   85%   119.2 BPM  3:26
   90%   126.2 BPM  3:14
   95%   133.2 BPM  3:04
  100%   140.2 BPM  2:55
wrote   D:\Drills\Zipp - Pour Quoi Royale - drill 70-100.mp3

14 file(s) in D:\Drills

Each rung gets a 4-beat click count-in at its own tempo, beat 1 accented, and a bar of silence after, so you settle into each speed before the song comes in. A rerun overwrites its own files: the drill's comment tag records the source, the ladder and the stems, and a second song with the same artist and title becomes ... (2).mp3. Every song is also a row in drills.csv next to the files: name, artist, title, the rung speeds, the length and the source path.

--ladder START-END:STEP	the speeds in percent (default 70-100:5); 50 <= start < end <= 100 and step 1 to 20
--passes N	repeat each rung N times (default 1)
--count-in N	clicks before each rung (default 4)
--gap-bars N	bars of silence after each rung (default 1)
--bars A-B	drill bars A to B from the beat grid instead of the whole song, e.g. --bars 9-16
--keep STEMS / --drop STEMS	drill only some stems, e.g. --drop vocals to practice with the band; separates each song again, so this needs a GPU and about one separation per song
--format mp3|wav	mp3 at 192 kbps CBR, resampled to 44.1 kHz, or wav at the source rate (default mp3)
--combined	one file for the whole setlist instead of one per song, songs separated by two 880 Hz beeps
--json	print a JSON manifest instead of the table

Below 65% the stretch leaves the 0.75x-1.5x range Signalsmith Stretch sounds best in, and the run says so before it renders anything. A WAV drill also carries a cue point on each rung's count-in, labelled with its speed (and on every repeat), and a combined WAV carries one cue per song, labelled with its title, so editors that read cues can jump straight to any speed or song. The --out folder is skipped when folders are scanned, so the drills never end up indexed and drilled themselves. Speed on an RTX 4090: the 14-track CC demo corpus became 6.1 hours of --drop vocals drills in about 10 minutes, 42 seconds per song, most of it separation. Stretching itself runs around 90x realtime on one core.

The count-in and gap timing assumes 4/4 at the indexed tempo, so on a song in another meter the gap is not exactly a bar; the whole song is held in memory while it stretches, so for a 20-minute mix prefer --bars; and a --keep/--drop drill is only as clean as Demucs, so listen once before trusting it.

What you can search for

--only names the stems that play alone, and everything else has to be silent. --no names the stems that have to be silent, and at least one of the others has to play. Stems are drums, bass, vocals and other (everything else: keys, guitars, pads, samples).

You want	Use
Drum breaks	--only drums (the default)
Acapellas	--only vocals
Bass and drums	--only bass,drums
Anything without drums	--no drums
Instrumentals	--no vocals

Other filters: --min-bars N (default 2), --bpm 90 or --bpm 85-100, --search "words" (every word has to appear in the artist, title, album or path), --cleaner-than -25 (only sections at least that clean), --sort clean|bars|bpm|artist, --limit N, --skip-seams, and --json for scripting.

How a bar is judged

For every bar, breakdig has the mean level in dB of each separated stem and of the mix.

• A stem is active in a bar if it is no more than 12 dB below the mix.
• A stem is silent if it is more than 30 dB below the loudest active stem in that bar, or below -90 dBFS. Silent wins over active, so a bar of digital silence has nothing active in it.
• A bar where a stem is neither is a grey zone and never matches. That is deliberate: it is what keeps a break with a quiet bassline under it out of "drums only".
• Consecutive matching bars become a section, and a section has to be at least --min-bars long.
• CLEAN is the loudest stem that should be silent minus the target stem, taken from the dirtiest bar in the section. -40 is very clean, -15 means you will probably hear something else in there. Results are sorted by it.
• BPM is 60 divided by the median beat interval inside the section, so tempo drift is handled per section. Bars come from detected downbeats, so 3/4 and changing meters work. The bar after the last downbeat counts too, if the music carries on for a full bar past it.

The 30 dB is --silence-db (and "Silence dB" in the UI). Demucs leaves 20 to 35 dB of bleed between stems on real records, so at 30 plenty of breaks you can clearly hear on the record will not match. Dropping it to 25 or 20 finds more at the cost of some bleed. The table's CLEAN column tells you how much.

Exports

Clips are cut from the original file, not from the separated stems, at the bar lines. Each edge is moved to the quietest point within 2 ms, which gets rid of most clicks; a loud bass note held across the bar line can still leave a faint one. Files are WAV at the source sample rate: 16-bit from 8- and 16-bit sources, 24-bit from everything else, including MP3 and AAC, which decode to more than 16 bits. A lossy file from a loud master often decodes past full scale, and those clips are written as 32-bit float so the transients are not clipped. They are tagged with artist, title with the bar range, album and BPM, and a comment holds the source path, times and content key. Exporting the same section again overwrites its file, even after the library has moved or the track was retagged. Names follow {artist} - {title} - {bars} bars - {bpm} BPM - {mm.ss}.wav, with characters Windows cannot handle replaced by _ and the name capped at 120 characters. If another section would get the same name, it becomes ... (2).wav.

Every clip also has a cue point on each beat, labelled beat 1, beat 2 and so on, in the WAV's standard cue chunk. Editors that read WAV cue points show them as markers to chop at. Plenty of software ignores them, which does no harm.

--keep drums (or the keep menu in the UI) exports only some stems instead of the mix. Each section is separated again with three seconds either side, so its edges come out clean, and the kept stems are put back at the source sample rate. That turns a near miss into a usable break: search --no vocals, look at the Stems column in the UI to see where the drums are, and export those rows with --keep drums. --keep drums,bass and the rest work the same way. The result is only as good as Demucs, so listen before you trust it. The name gets the stems before the bar count (... - drums - 4 bars - ...), and the isolated clip and the full-mix clip of the same bars are separate files.

--sp404 (or SP-404 MKII in the UI's format menu) writes 16-bit 48 kHz files named BRK_0001.WAV, BRK_0002.WAV and so on, plus an index.csv that maps each file back to its source. --sp404 sx (SP-404 SX / A in the UI) does the same at 44.1 kHz for the SX, the A and the original 404:

file,artist,title,bars,bpm,start,end,source,stems
BRK_0001.WAV,Bebop,Fidget Kaos,3,130.0,195.724,201.264,D:\Music\Breaks\SOSLP008\SOSLP008_01_BEBOP_DONT_BELIEVE_THE_HYPE-Fidget_Kaos.mp3,
BRK_0002.WAV,Innyu,For The Hell Of It,3,127.7,305.649,311.300,D:\Music\Breaks\SOSLP008\SOSLP008_02_INNYU_DONT_BELIEVE_THE_HYPE-For_The_Hell_Of_It.mp3,
BRK_0003.WAV,Innyu,For The Hell Of It,2,127.8,11.273,15.062,D:\Music\Breaks\SOSLP008\SOSLP008_02_INNYU_DONT_BELIEVE_THE_HYPE-For_The_Hell_Of_It.mp3,drums

Exporting into the same SP-404 folder again skips sections that are already there and numbers new ones after the highest number in the folder or the csv, so a number is never reused. Delete a BRK file and that section is exported again under a new number. Clips that would go past full scale are turned down to fit instead of being clipped. Editing index.csv in Excel is fine, including extra columns of your own and semicolon-separated saves, but close it before exporting again. A folder that already has clips at one rate refuses the other, so one card never ends up with both.

The web UI

breakdig ui serves on http://127.0.0.1:8765 and opens your browser. Pick stems, set the filters and press Find. The play button loads just that section and loops it with no gap at the loop point, so you hear how it will loop in a sampler. Space plays or stops the selected row and the arrow keys move through the list. Tick rows and press Export to write them to the folder at the bottom, then use "reveal" to open the file in Explorer. Ticked rows stay ticked when you sort or search again, so one export can collect sections from several searches; "clear" unticks them all, and an export unticks what it handled. The filters, volume and loop setting are remembered for next time.

The Stems column draws each bar of a section as four small cells, drums, bass, vocals and other from top to bottom, brighter the louder that stem is against the mix. The keep menu next to Export does what --keep does, for previews as well as exports, so you can hear the drums on their own before you write them out.

--port picks another port and --no-browser skips opening a tab. Previews are cached in the index folder, up to about 500 MB, and the oldest go first.

It only listens on localhost. --host 0.0.0.0 works but prints a warning, because anyone who can reach the port can browse your index and write exports anywhere you can.

Configuration

Setting	Default
--home or BREAKDIG_HOME	%LOCALAPPDATA%\breakdig	Index folder: index.sqlite, per-track profiles, index.log, the UI's clip cache
BREAKDIG_MODELS	%LOCALAPPDATA%\breakdig\models	Separation models, shared by every index
index --model	htdemucs_ft.yaml	Any audio-separator Demucs model
index --shifts	1	Demucs random shifts; 2 is slightly cleaner and twice as slow
index --retry-failed	off	Try files that failed last time again
index --exclude GLOB	none	Skip files and folders whose name matches, e.g. "*.removed.*" or stems. Repeatable

--home goes before or after the command: breakdig --home D:\breakdig-index find. For anything but index, a --home folder with no index in it is an error rather than an empty result, so a typo does not look like an empty index.

Speed

Measured on an RTX 4090 with the default model:

• 64 tracks, 2.5 hours of audio: 11.3 s per track, 12.6x realtime (breakdig stats).
• A 50-track folder, 1.8 hours of audio, indexed unattended: 9.4 s per separated track.
• A 21-minute MP3: 89 s.
• Exporting with --keep: the first clip takes about 9 s, most of it loading the model, and after that about 6x realtime. Eight sections, 7.5 minutes of audio in all, took 78 s.
• Re-running on an indexed folder: a second or two. Files whose size and modified time have not changed are not read at all, and neither are unchanged byte-for-byte copies.

The index is small: about 25 KB per minute of audio, so 90 KB or so for a typical track.

Files it handles and skips

MP3, FLAC, WAV, AIFF, M4A (AAC or ALAC), AAC, OGG Vorbis, Opus, WMA, WavPack, Monkey's Audio, Musepack and MP2. Mono, 8-bit, 24-bit, float and odd sample rates are fine. The same audio at two paths or in two formats is only separated once; the copy is recorded as a duplicate and takes over if the original is deleted. A lossless copy of a track first indexed from a lossy file takes over from it, so exports come from the better file. An original on a drive that is not plugged in counts as still there. The exception is a raw .aac file, which does not record its encoder delay, next to the same track in another format. A file that moves or is renamed keeps its index entry, and so does one you retag. One you trim or edit is analysed again.

The Recycle Bin, System Volume Information and macOS ._ files are not indexed.

Skipped, logged, and listed by breakdig stats --failures: files that will not decode, DRM-protected iTunes files (.m4p), and tracks with fewer than 8 downbeats (no_grid; mostly speech, ambient and very short files).

Vinyl rips with crackle can go through grooveclean first. Its batch mode writes a .removed difference file next to each cleaned side, which has enough of a beat to get indexed, so leave those out: breakdig index D:\Rips\cleaned --exclude "*.removed.*".

Files over 20 minutes are separated in 10-minute chunks. A section that crosses a chunk boundary is marked seam because the separation can glitch there; --skip-seams drops them.

Limitations

• On most commercial records the drums are almost never truly alone for two bars. Expect a handful of hits per album at the default setting, and use --silence-db 25 and --min-bars 1 to see near misses, or --keep drums to take the drums out of busier bars.
• Downbeat tracking can be wrong on rubato, very old recordings and music with no clear pulse. Bars far off a track's usual length are ignored rather than exported as nonsense.
• Separation quality is Demucs quality. Sections are judged on separated stems but cut from the original mix, so what you export is exactly what was on the record, unless you ask for --keep.
• Tested on Windows 11 with Python 3.11. It should run on Linux; macOS will only have the CPU, which is slow.

Development

py -3.11 -m venv .venv
.venv\Scripts\pip install -e .[test]
.venv\Scripts\python -m pytest -q

The gpu tests run real Demucs separation and are skipped without CUDA. Set BREAKDIG_MODELS to reuse a model folder between runs. CI runs everything else on Windows and Linux with the CPU build of torch.

Releasing

python -m build
python packaging\make_windows_zip.py
twine upload dist\breakdig-0.2.0-py3-none-any.whl dist\breakdig-0.2.0.tar.gz

make_windows_zip.py writes dist\breakdig-0.2.0-windows.zip for the GitHub release.

Credits

The demo uses Creative Commons music from the Internet Archive: "Retrovision" (MIXG032, CC BY-NC 3.0) with tracks by Zipp, Dog On Springs, Astat, Fedorov Mark, VAD and Igor Leontyev, and "Dont Believe The Hype" (SOSLP008, CC BY-NC-ND 2.5 IT) with tracks by Bebop and Innyu.

Separation is Demucs through audio-separator. Downbeats come from beat_this.

License

MIT