Published by PitchChanger · How these guides are made
How to use this workflow
Open the four-stem splitter
Use a browser with WebGPU. Starting separation loads or checks an approximately 181 MB model plus runtime assets. The splitter does not have a CPU fallback.
Separate a suitable recording
Choose a supported file up to six minutes and 250 MB, subject to memory and storage checks, then select Split four stems. Begin with a short passage containing the drum pattern you want.
Solo Drums and download
Listen to the Drums stem in the synchronized player and compare it with the other stems. Play your original file separately for a source comparison. Choose Drums in the export selector, prepare the WAV and download it.
Check more than the kick and snare
Listen for cymbal sustain, ghost notes, hand percussion and fills. Some of these sounds can leak into Music or Vocals, while a sharp guitar or vocal consonant may appear in Drums.
A pattern can remain recognizable even when the stem sounds imperfect in solo. Decide whether the timing and details you need are present. No numerical separation-quality score is promised for your file.
Use the stem as a practice reference
Estimate the tempo with BPM Finder or count it with Tap BPM, then compare the pattern against the metronome. For a difficult fill, trim a short section and slow it with the independent speed control.
The current model groups the drum kit and percussion into one stem. It does not separately export kick, snare, hi-hat and toms, create MIDI drum notes or provide a written transcription.
Drums and drumless are different downloads
Choose Drums to hear the estimated percussion. Choose Drumless to hear the remaining bass, music and vocals combined. Inspect the selected export label before downloading; these serve opposite practice needs.
The model input is 44.1 kHz stereo and the splitter exports WAV. If you convert to MP3 afterward, keep the WAV stem as your working copy. A processed stem does not by itself establish permission to redistribute the original song.
Check several kinds of drum detail before relying on the stem
Choose a few identifiable moments in your own recording and note their approximate positions. Start with a clear beat, then a fill and a quieter passage. Listen to the original separately, return to the same position in the stem player, and solo Drums. If a sound seems missing or misplaced, audition Music or Vocals at that position too.
Use the comparison to decide whether this particular stem serves your task. Following the main rhythm needs different detail from studying a soft ghost note. The worksheet below describes checks to perform; it is not a separation score or a prediction of results.
Swipe or scroll sideways to read all columns.
| Passage to choose | Listen for | Practical decision |
|---|---|---|
| A clear kick and snare pattern | Whether the main attacks remain easy to follow | Use it for rhythm study if those cues are clear. |
| A busy fill | Missing hits or other instruments that resemble hits | Keep the original nearby when learning the fill. |
| A cymbal decay or quiet ghost note | A cut-off tail or detail heard in another stem | Avoid treating the isolated stem as a complete kit transcription. |
Check the Drums download independently of the preview
Soloing Drums changes what the synchronized player auditions. The download uses the separate named mix selection. Select Drums in the download menu, choose Prepare WAV, and download the prepared file. Do not assume that a previous mute or solo combination determines the exported mix.
Open the saved WAV and listen to the same passages you checked in the preview. Confirm that it contains the expected drum material and includes the beginning and ending you need. The export is a 44.1 kHz stereo, 16-bit WAV; it does not provide separate kick, snare, and cymbal channels. Splitting kit pieces or adjusting their individual levels requires another audio workflow.
Keep the original and the checked WAV before trimming or slowing a practice passage. If you make an MP3 copy for sharing, check that copy as well and retain the WAV for further edits.
Use a timing grid to investigate a quiet hit
For an illustrative steady 120 BPM passage in 4/4, one quarter-note beat is 0.5 seconds, an eighth note 0.25 seconds and a sixteenth note 0.125 seconds. Choose a clear downbeat in the original and use these offsets as a listening guide for nearby quiet hits. A grid gives expected positions; it does not prove that a ghost note exists or that a drummer played mechanically on the grid.
Solo Drums around one suspected hit, then audition Music and Vocals at the same position if the detail seems absent. Keep the original nearby as a separate comparison. Noise or spill at an expected offset should not be turned into a transcribed hit solely because the arithmetic places it there.
Swipe or scroll sideways to read all columns.
| Offset from a chosen beat at 120 BPM | Subdivision position |
|---|---|
| 0.000 s | Beat |
| 0.125 s | Second sixteenth |
| 0.250 s | Eighth-note midpoint |
| 0.375 s | Fourth sixteenth |
| 0.500 s | Next beat |
Slow a short drum passage while keeping its place
Download the Drums WAV, trim a passage that includes a clear beat before and after the detail, then load that WAV into the independent speed tool. At 50% speed, a 0.125-second gap becomes 0.250 seconds and a 120 BPM passage becomes 60 BPM. These are calculated changes; the speed tool does not label drum hits or generate MIDI.
Check the quiet attack in both the original-rate and slowed files. Slowing can expose timing but cannot recover percussion missing from the separation. Retain the untrimmed stem and note the excerpt’s source start externally. If your task requires exact kit-piece identification or notation, use the original as well as the stem rather than treating the separated waveform as a complete transcription.
Understand what Music · Other instruments groups together
The four-stem model estimates Drums, Bass, Vocals and Other. The interface labels the last category Music · Other instruments. That category can include guitar, piano, keyboards and other material that is not assigned to the three named groups. It is a broad estimated part of the mix, not a separate guitar or piano channel.
For an illustrative arrangement with a drum kit, bass guitar, lead vocal, rhythm guitar and piano, the expected categories are shown below. This is an arrangement chart explaining the model’s output vocabulary, not an annotated multitrack test or a report of where a particular recording’s sounds were measured. Actual assignment can include bleed and missing detail.
Swipe or scroll sideways to read all columns.
| Illustrative arrangement element | Expected broad category | Isolation limit |
|---|---|---|
| Drum kit and percussion | Drums | Kit pieces are grouped |
| Bass guitar | Bass | Other low-frequency sounds may interfere |
| Lead vocal | Vocals | Backing voices and spill can complicate the estimate |
| Rhythm guitar and piano | Music · Other instruments | Guitar and piano are not separate outputs |
Choose the export that matches the part you need
To investigate a guitar-like sound contaminating Drums, audition Music at the same position and compare the original separately. If your goal is percussion study, select the named Drums export. If you want the song without its estimated percussion, select Drumless; that mix combines Bass, Music and Vocals. Solo controls affect preview and do not define an arbitrary download mix.
If you need only the guitar while removing piano, the four-stem categories cannot supply that instrument-specific export. Use a source multitrack or a workflow that explicitly supports the needed separation, and check the result. Repeatedly selecting Other does not create additional instrument channels. Keep the original and listen to the saved named WAV so your practice decision follows the deliverable rather than the category label.
Keep the four downloaded stems on one timeline
If you want to edit the separated parts in another application, export Drums, Bass, Vocals and Music from the same separation result. Choose each named output and prepare its WAV separately. Label the files with one shared source identifier. Do not trim the leading silence from individual stems before importing: a quiet opening can still occupy the time needed for the parts to line up.
In a DAW, place all four files at the same timeline start and check their displayed durations. Disable any automatic tempo matching or clip stretching for this comparison. For an invented eight-second source whose first drum attack occurs at 1.25 seconds, removing one second from only Drums makes that attack occur at 0.25 seconds relative to the other untrimmed files. Equal filenames or tempo labels cannot correct that one-second offset.
Listen at the opening and near the ending after import. An equal starting position with a later mismatch suggests checking per-clip stretching, rate interpretation or different source versions. If every part needs the same trim, group the clips and apply the edit together in the receiving editor. Keep the exported originals. Recombining model estimates is useful for checking alignment, but it is not proof of a bit-identical reconstruction of the original recording. This guide provides a timeline protocol rather than a measured DAW session.
Swipe or scroll sideways to read all columns.
| Alignment check | Expected condition | If it fails |
|---|---|---|
| Clip starts | Same timeline position | Restore a shared start |
| Clip timing | No individual stretching | Inspect import and tempo settings |
| Later cue | Parts remain aligned | Check rates, edits and source versions |
Common questions
Can I get a perfectly clean isolated drum kit?
Not reliably from every finished mix. Separation can retain bleed, lose percussion detail or create artifacts where instruments overlap.
Can I isolate drums with the two-output vocal remover?
That page separates an estimated instrumental and vocal output. Use the four-stem splitter for a dedicated Drums category.