Stem Splitter

Short answer: the Stem Splitter plugin creates separated audio parts (vocals, drums, bass, guitar, and more) from a song's full mix, and can transcribe or re-time its lyrics — all from inside FeedBack, without a terminal.

The Stems and Stem Mixer plugin plays and balances stems. Stem Splitter is the one that makes them.

When To Use This

A song only has a full mix, and you want isolated parts to practice against.
A song is already split, but you want to split it again — a newer model gives cleaner results, or an earlier conversion came out faulty.
You want to remove the part you play — mute the guitar and play it yourself over the rest of the band, or drop the vocals for a karaoke-style backing track.
You want to isolate one part instead — solo the drums to hear the groove, or the vocals to learn a line.
A song has no lyrics and you want them transcribed from the vocals.
A song's lyrics are correct but their timing drifts, and you want them re-aligned.

Before You Start

Splitting and transcribing are heavy audio jobs. They can run three ways, and the plugin picks whichever you've set up:

On this machine — download a local engine once, then everything runs offline. A GPU is faster, but a CPU works.
On a server — point the plugin at a Demucs/WhisperX server (yours or a shared one). See Demucs Server.
The built-in managed server — the plugin can install and run a local server for you (Settings → Stem Splitter → Install server + models). This is the easiest path and needs no command line.

The first job of each kind downloads its models — up to a couple of gigabytes, once. Nothing large is ever downloaded until you ask for it: the plugin warns and waits for your confirmation before any big download.

First-Time Setup, In Order

If you've never split a song before, do this once. After that, splitting is a single click from a song's menu.

Open Settings → Stem Splitter.
Pick how it runs. The easiest choice is the built-in managed server: click Install server + models. (Alternatively, download a local engine to run fully offline, or enter a server URL if you already have one.)
Wait for the download. This is the one large download — up to a couple of gigabytes. It happens once, and you can watch its progress on this page.
Confirm it's ready. Click Test status. It should report the server is up and show its device (CPU or GPU).
Now split a song (below). The models are already downloaded, so the job starts right away.

You can also skip straight to step 5: if the models aren't ready when you split, the plugin tells you the size, asks first, downloads them, and then runs the job automatically. Either order works — this section is just the calmer path.

Steps

Split a song into stems

In your library, open the menu on a song's card.
Choose Split stems. (It's greyed out on songs that are already fully split.)
If models aren't downloaded yet, the plugin tells you the size and asks first. Approve it.
Watch progress in the queue — click the "Split queued" toast, or open Settings → Stem Splitter → Open the queue.
When it finishes, open the Stem Mixer to hear the parts.

The original full mix is always kept in the song, so a split never throws audio away. And if you've added your own stems to a song — a click track, a backing part — a split leaves them exactly where they are.

Re-split a song that already has stems

Use this to redo a split with a better model or engine, or to fix a faulty one.

Open the menu on the song's card.
Choose Re-split stems…. (It appears only on songs that already have stems.)
A picker lists the song's replaceable stems, all checked. Uncheck any stem you've replaced yourself and want to keep — an unchecked stem isn't touched at all: same file, same bytes.
Click Re-split and watch the queue, as with a normal split.

Only the checked stems are overwritten. Stems the splitter can't produce — a click track you added, a custom backing stem — and the original full mix are never touched, and the picker says so. If you replaced a song's guitar with your own recording, uncheck guitar and it survives every re-split.

Transcribe lyrics

Open a song's card menu.
Choose Transcribe lyrics. (Enabled only when the song has no lyrics yet.)
Approve the model download if prompted, then watch the queue.

Transcription listens to the vocals and writes the words and their timing. An instrumental produces no lyrics — that's a normal result, not an error.

Re-align existing lyrics

Use this when the words are right but the timing is off — lyrics pasted from the web, imported from a Guitar Pro file, or transcribed against a different mix.

Open a song's card menu.
Choose Re-align lyrics to vocals. (Enabled only when the song has both lyrics and a vocal stem.)

Re-align keeps your words exactly as they are and only fixes their timing. It never replaces a word — if you want the words themselves redone, use Transcribe lyrics instead. Re-align needs a server (the built-in managed one is fine); the local engine can transcribe but can't re-align.

What You Should See

Card-menu actions: Split stems, Re-split stems…, Transcribe lyrics, Re-align lyrics to vocals, each enabled only when it applies to that song.
A queue you can reach from the toast or from the plugin's Settings page, showing each job's progress and a badge with the number running.
A failed job shows its full error, with a Copy button for pasting into a bug report.

Screenshot placeholders:

SCREENSHOT: song card menu showing the three Stem Splitter actions
SCREENSHOT: Stem Splitter settings — engine choice, managed server card, Open the queue button
SCREENSHOT: the queue with a split and a transcription running
SCREENSHOT: the re-split picker with one stem unchecked

Managing the Built-In Server

Everything below lives in Settings → Stem Splitter, and needs no terminal.

Button	What it does
Install server + models	Sets up the managed local server and downloads its models.
Start / Stop	Runs or stops the local server. Auto-start is on by default once installed.
Test status	Checks the server is up and reports its device (CPU/GPU) and warmup state.
Prepare models	Downloads the model weights on their own, ahead of your first job.
Check for update	Refreshes the server to pick up fixes, without re-downloading models.

The status chips distinguish loading (reading a model you already have from disk) from downloading (fetching over the network), so a quick warm-up isn't mistaken for a fresh multi-gigabyte pull.

Running FeedBack in Docker

If you run the FeedBack app itself in a container, the plugin can't install a server on your host. Instead it offers a compose snippet to run the server as a companion container (always shown, recommended), and — only when the Docker socket is already available to the app — a one-click option. GPU is requested only when your Docker setup actually advertises one.

Caveats and Limitations

Separation is powerful, but it isn't magic. Knowing these up front saves disappointment:

Stems are split by instrument type, not by individual part. A song with a lead and a rhythm guitar produces one guitar stem containing both — the model can't tell two guitars apart. The same goes for two vocalists, a doubled bass, and so on.
Separation is not lossless. Isolated stems carry some bleed from other instruments, and re-mixing every stem does not perfectly reconstruct the original. The full mix is always kept in the song for exactly this reason — use it when you need the untouched audio.
Quality depends on the source and the model. A dense or heavily-processed mix separates less cleanly than a sparse one. Different models trade speed for quality.
The stems a model makes depend on the model. A 4-stem model gives vocals, drums, bass, and other; a 6-stem model adds guitar and piano. Parts the chosen model doesn't isolate fall into other.

Common Problems

Problem	Try This
The action is greyed out	Split needs an un-split song; Re-split needs a song that already has stems; Transcribe needs a song with no lyrics; Re-align needs both lyrics and a vocal stem.
"Re-align needs a server"	Re-align can't run on the local engine. Configure a server, or use the built-in managed one.
A job seems stuck at the start	It may be downloading models. Check the queue and the Settings page for download progress.
Models re-download every launch	Update the server with Check for update — an older server had a bug that deleted weights.
Transcription returns nothing	The vocal stem may be silent (an instrumental) or in a different language than expected.
A job failed	Open the queue, read the full error, and use Copy if you need to report it.
The whole batch failed at once	Often a server that's unreachable or missing its models. Check Test status.

Applies To

Version: Stem Splitter 0.5.x Platforms: Windows, macOS, Linux (managed server and local engine); any host for a remote/Docker server