Separate stems

The Separate page splits a song into stems, such as vocals, drums, bass, and the rest, with a model you choose. The stems land on new tracks under the song.

It needs the Stem separation component: about 10 MB, plus 200 MB on Apple silicon, plus each model you download. If the page shows Install, click it, or see Add the tools you need.

Separate a song

1. Select one or more audio items. Each is read as it plays, with its trims, rate, pitch, stretch markers, and take FX.
2. On the Separate page, make sure Model is chosen at the top, not Workflow.
3. Choose a model. The list shows your favorites and the models you have downloaded; Model library (the box button at the top) has the rest. See Choose a model.
4. Under Writes, tick the stems to keep. All and None tick or untick them all.
5. Click Separate, or Separate 3 for three items. If the model isn’t downloaded yet, the button reads Get with the model’s size, and separate: the model downloads first.

Expected result: a progress bar at the bottom of the page, with Cancel. Then, for each item:

• Each stem is a WAV file (32-bit float) named after the song and the stem, such as song_vocals.wav, written beside the file the item plays.
• Each stem is on a new track under the item’s track, named after its file and lined up with the item, in one undo step.

If a stem’s file is already there, PresTools asks Replace this file? (or Replace these files?) and lists them. Replace writes over them; Keep them stops.

Add summed tracks

Two boxes under the stem list add tracks that sum stems for you:

• Additional track with selected stems mixes the ticked stems into one more file and track, such as song_bass+drums.wav. It needs two or more stems ticked.
• Additional track with deselected stems writes the song minus the ticked stems, song_rest.wav. Tick the vocals, for example, to get a backing track without them.

The boxes show once the model is downloaded, for models that separate two or more stems of their own. A stem marked derived is the mix minus the model’s other stem, not separated on its own; such models have no summed tracks.

Separate files that are not in the project

With no item selected, click Choose files or a folder at the bottom of the page. The file chooser offers WAV, MP3, FLAC, M4A, AAC, OGG, and Opus files. A folder gives the audio files in it, AIFF included, but not those in its subfolders.

• The stems go beside each file, and onto new tracks at the bottom of the project in front, at the edit cursor.
• Tick Multi-project to give each file a project of its own instead. PresTools opens each in a new tab, with the song and its stems, saves it beside the song as song.rpp, and closes it.

Forget this file (the × beside the file name) clears the files you chose. Selecting an item replaces them.

Choose a model

The Model list on the Separate page shows your favorites and the models you have downloaded. The Model library button opens the whole catalog:

• Search by name, architecture, or stem, and filter by Category, such as Vocals, Whole song, or Karaoke.
• Show installed models only (the check button) hides the models you haven’t downloaded.
• The star adds a model to your favorites. Install (the download button) downloads it; Delete (the bin) deletes it.
• Double-click a model to separate with it.

Each model’s size shows beside it. Models are downloaded one at a time, from each model’s own source, into PresTools’ models folder. The catalog does not list licenses: check the terms of any model you use commercially.

Model settings (the sliders button beside the star) opens the model’s own settings, kept for that model.

A model stays loaded between runs, so the next run starts sooner. To free the memory, open Settings, find Held in memory under Models, and click Release.

In Prest and PresTools’ next release

The newer PresTools built into Prest, which comes to REAPER with PresTools’ next release, runs BS-RoFormer and Mel-Band RoFormer models only. Its library shows each model’s license, and hides models without a stated license until you choose to show them. A 0.1.0 workflow that uses another kind of model, such as Demucs, does not run there.

Workflows

A workflow runs several steps in a row: for example, split the song, then clean up its vocal stem, then map the tempo and find the chords. Use one when you prepare many songs the same way.

Build a workflow

1. On the Separate page, choose Workflow at the top.

2. Click New workflow (the plus button) when you have none, or Edit this workflow (the sliders button). The workflow editor opens; New workflow there starts another.

3. Name the workflow.

4. Click Add a step or a node, and choose:

   • Separation step: a model that splits audio into stems.
   • Tempo map: finds the beats and draws the tempo map, with a count-in.
   • Detect chords: writes what is played onto the Chordtrack.
   • Recognize lyrics: writes down what is sung, to a text file to correct before aligning.
   • Align lyrics: times words from a text file to the singing, onto the Lyrics track.

5. Drag from the song, or from a step’s stem, to the next step or node. A step reads exactly one thing, so a new cable replaces the one it lands on. Double-click a cable to remove it. A separation step shows its stems once you choose its model.

6. Drag each stem you want as a file to Output, and name the file there. Wildcards lists what a name can hold; click one to put it in the name you last clicked into:

   Wildcard	Becomes
   %filename%	The song’s file name.
   %stem%	The stem, as the model names it.
   %model%	The model.
   %index%	The file’s number among those the workflow writes: 01, 02, and so on.
   %input_number%	Always 01: each song runs on its own.
   %yyyyMMdd%	The date.
   %hhmmss%	The time.
   %ddmmss%	Day, minute, and second.

   A name with no wildcard is added to the song’s name: vocals writes song_vocals.wav. An empty name is %filename%_%stem%_%model%.

7. Click Save.

This step’s settings opens a step’s settings. The Tempo map step has the detection settings and Count-in in bar 1. Detect chords has Snap to the grid, which needs bars: put a Tempo map step before it, or map the tempo first. Recognize lyrics has Write to (default %filename%.txt, beside the song), Whisper model, and Language. Align lyrics reads its words from Words from, the same default, so it follows Recognize lyrics, and has Language.

Workflows are saved as .yaml files in pymss’s workflow format, so Import a workflow file takes a pymss workflow written elsewhere (.yaml or .yml). Open another workflow switches between them; Delete this workflow deletes its file.

Run a workflow

1. Select the items, as for a model.
2. On the Separate page, choose Workflow, then the workflow.
3. Under Writes, tick the outputs to write. An unticked output isn’t written, and a step that is then left with nothing to do doesn’t run.
4. Click Run.

Missing models download first. Stems land on new tracks under the item, as for a model. The tempo, chord, and lyrics steps write into the item’s project, where the song sits.

Those steps need a project: for files you chose by hand, tick Multi-project, or the run stops with song.wav is not in a project; tick Multi-project to give each file one.

What it runs on

The GPU, unless you turn it off in Settings: MLX or Metal on Apple silicon, CUDA or ROCm on NVIDIA and AMD GPUs. If the GPU can’t run a model, the run finishes on the CPU, and that model goes straight to the CPU for the rest of the session.

Next: Map the tempo.