Split a song into stems

Stem separation splits a finished mix into its parts: vocals, drums, bass, and everything else ("other"). Use it to make a karaoke or instrumental version, pull a vocal for a remix, practise along with the drums muted, or clean up one part. Outputs below are trimmed copies of what ferraus 0.1.0 printed for a 21-second test song.

You need: a song (song.wav; MP3, FLAC, and M4A work too), about 170 MB of disk space for the model, and a network connection the first time.

Which tier to use

Separation uses neural-network models from the ferraus model registry, and every tier has a verified download. The default tier, auto, picks quality (a Mel-Band RoFormer for the vocals plus fine-tuned Demucs for drums, bass, and other: about 960 MB of models, about 7 minutes for a 4-minute song on a laptop CPU), which gives the cleanest vocals. --tier balanced (four fine-tuned Demucs models, about 660 MB, about 3 minutes) is the middle ground. The commands below use --tier fast (Demucs v4 htdemucs, 166 MB, MIT license, under a minute for a 4-minute song) to keep the first run quick. See Stems and separation tiers for all tiers and their measured quality.

1. See what would happen

ferraus separate song.wav --tier fast --dry-run --json

{"ok": true, "command": "separate", "dry_run": true,
 "result": {"plan": [{"model": "separate/htdemucs@v4-onnx.2026-05-21", "weight": 1.0, "cached": false}],
            "outputs": [{"stem": "vocals", "path": "song.vocals.wav"}, {"stem": "drums", "path": "song.drums.wav"},
                        {"stem": "bass", "path": "song.bass.wav"}, {"stem": "other", "path": "song.other.wav"}]}}

Without -y ferraus never downloads anything; it stops with exit code 5 and prints the exact command that fetches the model:

{"ok": false, "error": {"code": "model_not_cached",
  "message": "model is not cached and downloads are disabled: separate/htdemucs@v4-onnx.2026-05-21",
  "hint": "ferraus models pull separate/htdemucs@v4-onnx.2026-05-21", "exit_code": 5}}

2. Separate

ferraus separate song.wav --tier fast -y --json

warning: downloading separate/htdemucs@v4-onnx.2026-05-21 (165.6 MB, license MIT)

{"ok": true, "command": "separate",
 "outputs": [{"path": "song.vocals.wav", "role": "vocals"}, {"path": "song.drums.wav", "role": "drums"},
             {"path": "song.bass.wav", "role": "bass"}, {"path": "song.other.wav", "role": "other"}],
 "models": [{"name": "separate/htdemucs", "version": "v4-onnx.2026-05-21", "sha256": "d05c269d..."}],
 "metrics": {"quality_vocals": 0.018, "quality_drums": 0.901, "quality_bass": 0.516, "quality_other": 0.869},
 "result": {"sum_to_mix": {"applied": true, "target": "other", "residual_peak_db": -15.3, "clipped": 0}, ...}}

The model is cached after the first run (ferraus models list --cached), and later runs work offline. On an Apple-silicon laptop a 21-second song takes about 5 seconds on the CPU.

Listen to each stem:

  • vocals: the voice alone. Expect some cymbal "wash" and reverb tails from other parts (bleed); a little is normal.
  • drums: kick, snare, and cymbals. Listen for vocal "s" sounds leaking in.
  • bass: should be solid in the low end; very low synth parts sometimes land in other.
  • other: everything else (guitars, keys, pads, synth leads).

The quality_* numbers are the model's own confidence per stem (0 to 1). Our test song is a synthesized instrumental, so quality_vocals is 0.018: there is no voice, and the lead synth went to other. A low score on a stem you know is present means that stem needs a careful listen.

3. The stems add back to the song

With --sum-to-mix on (the default) whatever the model missed goes into other, so the four files add back to the original exactly. You can check that:

ferraus mix song.vocals.wav song.drums.wav song.bass.wav song.other.wav -o sum.wav
ferraus compare song.wav sum.wav

{"rms_diff_db": -240.0, "si_sdr_db": 200.0, "identical": true, "verdict": "identical"}

4. Common variations

ferraus separate song.wav --tier fast --stems 2 --outdir two         # vocals + instrumental
ferraus separate song.wav --tier fast --instrumental --outdir karaoke # one file with the vocals removed
ferraus separate song.wav --tier fast --only drums,bass --outdir rhythm
ferraus separate song.wav --tier fast --stems 6 --outdir six         # adds guitar and piano
ferraus separate song.wav --tier fast --drum-split --outdir kit      # drums → kick, snare, toms, cymbals

two/song.vocals.wav  two/song.instrumental.wav
karaoke/song.instrumental.wav
rhythm/song.drums.wav  rhythm/song.bass.wav
six/song.vocals.wav  six/song.drums.wav  six/song.bass.wav  six/song.guitar.wav  six/song.piano.wav  six/song.other.wav
kit/song.vocals.wav ... kit/song.kick.wav  kit/song.snare.wav  kit/song.cymbals.wav  kit/song.toms.wav

--stems 2 --tier quality runs only the RoFormer vocal model, so a clean karaoke track costs about two thirds of a full quality run. --drum-split downloads the DrumSep model (335 MB) on first use; it has no separate hi-hat output (hi-hats land in cymbals), and --dry-run lists the extra outputs before anything runs.

--outdir keeps each run's files apart; without it, a second run stops with output_exists instead of overwriting (pass --overwrite to replace).

5. Long songs, agents, and batches

Separation is the slowest command in ferraus. Memory does not grow with the song: models run one at a time over the song window by window, the song is decoded and the stems are written block by block, and an hour-long input spills its intermediate signals to temporary files (removed afterwards) so it stays under 4 GB (measured on a laptop: 1.5 GB for fast, 2.8 GB for quality). A larger --max-memory keeps everything in memory, which on a laptop CPU is no faster. Run long jobs in the background and collect the result later:

ferraus separate song.wav --tier fast --outdir stems --async --json

{"ok": true, "job_id": "1792400395782-593d"}

ferraus jobs status 1792400395782-593d   # state, percent
ferraus jobs wait 1792400395782-593d     # blocks, then prints the separate result

What next

Reference

separate · models list · models pull · compare · jobs wait