Stems and separation tiers

ferraus separate splits a mix into stems: 2 (vocals, instrumental), 4 (vocals, drums, bass, other), or 6 (adds guitar and piano). The models are neural networks from the ferraus model registry, run through ONNX Runtime on the CPU or a GPU.

Tiers

A tier is a named recipe of registry models. This table is what ferraus separate --dry-run plans and ferraus models list --task separate reports:

Tier	Models (4 stems)	Download
`fast`	Demucs v4 `htdemucs` (one model for all four stems)	166 MB, MIT
`balanced`	Demucs v4 fine-tuned specialists `htdemucs-ft-drums`, `-bass`, `-other`, `-vocals` (one per stem)	4 × 166 MB, MIT
`quality`	Mel-Band RoFormer `mel-roformer-kj` (Kimberley Jensen's vocal model) for vocals, plus the fine-tuned Demucs drums, bass, and other	464 MB + 3 × 166 MB, MIT
`max`	weighted ensemble: RoFormer 0.8 + `htdemucs-ft-vocals` 0.2 for vocals; fine-tuned Demucs 0.6 + `htdemucs` 0.4 for drums, bass, and other; 2 test-time-augmentation shifts, each also channel-swapped	all of the above, MIT

Every model in every tier has a recorded SHA-256 and a pinned download URL, so all four tiers work. The RoFormer's training data is not disclosed by its author (its weights are MIT); ferraus models list shows each entry's license note.

--stems 6 uses htdemucs-6s in every tier (136 MB, MIT). Guitar and piano are best-effort quality.
--stems 2 runs only the models that produce vocals; the instrumental is the mix minus the vocals (exactly, with --sum-to-mix on).
--model <name or .onnx file> pins one model instead of a tier; --ensemble a,b --weights 0.6,0.4 builds your own ensemble from registry models.
Several registry entries (bs-roformer-viperx-1297, bs-roformer-sw, scnet-xl, mdx-kim-vocal-2, mdx-kara-2) are listed with the flag no_clear_license and belong to no tier; --model with --accept-license uses one explicitly.

The default tier is auto: the best tier whose models are all verified and that the registry does not hold back (a tier's auto_skip field), tried in the order quality → balanced → fast. With this release's registry auto means quality, and the result records "tier": "quality (auto)". A registry whose better tiers lack verified models falls back to balanced or fast and says so in a tier_fallback warning (ADR 0007).

Quality

Separation quality is measured as SDR (signal-to-distortion ratio, in dB; higher is better). Published Multisong results put fine-tuned Demucs v4 at about 8.4 dB on vocals and Kimberley Jensen's Mel-Band RoFormer at about 11 dB. On the first 12 MUSDB18 test excerpts (the 7-second CC BY 4.0 set; the RoFormer's training data is unknown, so it may have heard them) ferraus measured the median SI-SDR per stem, in dB:

Tier	Vocals	Drums	Bass	Other
`fast`	8.6	7.9	7.8	3.3
`balanced`	9.0	7.7	8.1	3.3
`quality`	11.9	7.7	8.1	4.8
`max`	12.1	8.1	8.2	4.8

SI-SDR is scale-invariant SDR against the true stems; MUSDB18's published SDR figures use a different metric (BSSEval), so compare tiers with each other rather than with papers. The quality tier beats balanced on vocals by 2.9 dB here; max adds about 0.2 dB on vocals and up to 0.4 dB elsewhere for about 5 times the work.

The fast tier is good enough for practice tracks, karaoke, and remix sketches; expect some bleed (cymbals in the vocal, reverb tails in the wrong stem). quality gives clearly cleaner vocals; drums, bass, and other come from the same fine-tuned Demucs models as balanced.

Each stem's JSON result carries a quality estimate between 0 and 1 (mask confidence for a single model, agreement between models for an ensemble). It is not calibrated to SDR, but a low value flags a stem to check by ear. A stem that is absent from the mix (no vocals in an instrumental) scores near 0.

Speed and memory

Measured on an M1 Pro laptop's CPU (10 cores) with a 4-minute song, at the default --max-memory 2GB (details and method in docs/dev/PERFORMANCE.md):

Tier	Time	Peak memory
`fast`	47 s	1.9 GB
`--stems 6`	43 s	1.5 GB
`balanced`	3.0 min	1.9 GB
`quality`	7.0 min	3.0 GB
`max`	34 min (4 passes of 6 models)	2.9 GB

Separation runs the models one at a time and streams each over the song window by window; the input is decoded straight into that stream and every stem is encoded block by block, so memory does not grow with the song: a 60-minute input stays under 4 GB in every tier (measured: fast 1.5 GB, balanced 1.6 GB, quality 2.8 GB). Long inputs spill their intermediate signals to temporary files (removed afterwards, also on error or cancel). When a model's own working memory exceeds --max-memory, ferraus loads it in a leaner mode (on a laptop CPU it is as fast as the full one); result.memory reports the plan (spilled, lean_sessions, fits). The quality tier's RoFormer needs about 2.8 GB on its own, so that tier goes over a 2 GB budget and says so with "fits": false. A supported GPU (ferraus accelerate install, see Troubleshooting) helps most for quality and max.

Options that matter

--sum-to-mix on (default) assigns whatever the models missed to other, so the stems add back to the original exactly: ferraus mix of the four stems against the input gives "verdict": "identical" in ferraus compare.
--only vocals,drums writes only those stems; --instrumental writes one file without the vocals.
--drum-split splits the drum stem into kick, snare, toms, and cymbals with DrumSep (a hybrid Demucs fine-tune, MIT; hi-hats land in cymbals). --extended guitar,piano takes guitar and piano out of other with htdemucs-6s. Each cascade output carries a quality flag, and --dry-run lists the cascade stages with their models and outputs.
--vocal-split splits the vocal stem into song.lead.wav and song.backing.wav. No lead/backing model has a clear license, so the default is built in (separate/melody-mask, nothing to download): it tracks the lead's pitch, keeps the energy on the lead's harmonics, sends other voices' harmonics to the backing, weights stereo input toward the centre, and makes backing = vocals − lead, so the two add back to the vocal stem. It is best effort ("method": "melody_mask"): it cannot pull apart a harmony in unison or at the octave, it takes a backing part that sings alone for the lead, and it warns when the pitch track is unreliable (vocal_split_pitch_unreliable) or the stem has little backing (vocal_split_little_backing). On synthetic test mixes it improved the lead by about 1 dB SI-SDR on mono input and 3–4 dB on stereo input with panned backing. --vocal-split-tracker crepe|swiftf0 uses a model pitch tracker instead of pYIN. --cascade-models vocal-split=<model> runs a separation model instead: an ONNX file of your own, or the UVR karaoke model separate/mdx-kara-2 (no license stated for its weights; never used by default; it needs --accept-license once and warns on every use).
--overlap 25% and --segment auto control the processing windows; --shifts 2 adds test-time augmentation (slower, slightly cleaner).
Models download on first use with -y, or pull them first: ferraus models pull --all-for separate --tier quality. ferraus models list --cached shows what is on disk.

See the tutorial and the remix recipe.