악기 연습을 위한 스템 분리 활용법

Why stem separation helps practice

Two difficulties come up most often when learning an instrument: it is hard to pick out exactly what your part sounds like inside a finished mix, and practicing alone means no band to play with. Stem separation splits a mixed recording into tracks (stems) such as vocals, drums, bass and everything else. With them you can solo your part to transcribe it, or mute it and play along with the rest. That once required multitrack masters or purpose-made practice tracks.

It is not magic, though. The output has limits, and knowing them tells you what to use it for and what not to. This guide starts with principles and routines that apply to any tool, and near the end shows one example, an app I build.

How separation works

The sound from a speaker is already one waveform with every instrument added together. Your ear can pick instruments out of it, but for a computer, splitting "this is bass, this is vocals" from the waveform alone is hard. Modern separation models use neural networks trained on many songs, learning pairs of mixed audio and its original parts, then estimating each part from the mix of a new song.

The Demucs family of open-source models is described in a research paper that combines a hybrid time-and-spectrogram design with transformers. That paper (Hybrid Transformers for Music Source Separation) reports an SDR of 9.20 dB on MUSDB when trained with 800 extra songs [1]. SDR measures how close a separation is to the original, higher being better, but a single number does not capture how it sounds. The model splits into four stems by default (drums, bass, vocals, other) and offers an experimental six-stem model adding guitar and piano, whose documentation itself says the piano stem is "not working great at the moment" [2].

The key point is that separation is an estimate, not an extraction of the truth. The model guesses which parts of the mix belong to each instrument. It does well on arrangements it saw often in training, such as vocals plus a band, and can go wrong on unusual sounds or where many instruments overlap.

Quality limits: artifacts and bleed

• Artifacts: separated tracks can sound watery, metallic or warbly because the model's estimate does not match the original exactly.
• Bleed: another part leaks in, such as vocals faintly audible in the drum stem or hi-hat in the vocals. It is worse where sounds share the same frequency range.
• Song dependence: sparse arrangements with clean recording separate well; songs full of backing vocals, effects and reverb separate worse.
• Finer parts: guitar and piano are usually less accurate than the basic four. The Demucs README says guitar is decent while piano has much bleed and artifacts [2].
• Not a perfect sum: adding all stems back together does not give the original exactly.

The practical rule follows: treat the result as practice material and do the final check of a transcription against the original, the score and your own ear. If one note in a separated guitar track sounds doubtful, go back to the original and compare the same bars.

Input tips for better results

• Use lossless or high-bitrate audio. Heavily compressed sources have already lost information, so noise in the separation stands out more. Start from WAV, FLAC or ALAC if you can.
• Studio recordings beat live ones. Audience noise and room sound blur the boundaries between parts.
• Check on headphones. On speakers it is hard to tell bleed and artifacts apart.
• Decide what to mute by your goal (table below).
• Keep output formats straight. For practice alone, muting and unmuting inside the player is enough; exporting files is only needed to use them in another player or recording program.

Workflows by instrument

Drummers

Mute the drums and play along with the real bass, guitar and vocals. That builds a feel for the song's flow and bar structure better than a bare metronome. Soloing the drum track shows fill order and hi-hat patterns. For a hard fill, loop one or two bars, slow down, and sort out the sticking first.

Bassists

Bass sits low, so ordinary earbuds make pitches hard to tell apart. Solo the bass, slow it down if needed and check each note. After transcribing, mute the bass and play with the drums to lock in with the kick.

Guitarists and keyboardists

Guitar and piano are the least consistent stems because many instruments overlap in their range. Treat the separated track as a clue, not the answer, and alternate with the original where unsure. Once you know the chord progression, muting the guitar and playing with the rhythm section is the most practical routine.

Singers

Soloing the vocal reveals breath placement, pitch handling and pronunciation. Sing along on a loop until it feels familiar, then mute the vocal and sing over the backing. Faint reverb from the original vocal can remain in the backing, which at first may blur with your own voice.

Starting slow and building speed

As important as separation is pitch-preserving speed change plus loop repeat. If pitch drops with speed you can no longer play along in tune, so use time-stretching that keeps pitch. At very slow speeds the sound may smear slightly, so alternate with the original to keep your sense of the real tempo. A suggested routine:

1. Mark one or two hard bars as an A-B loop. Keep it short (one to four bars).
2. Start at about 60 percent of the original speed and check that you can play it perfectly.
3. After three clean repetitions in a row, raise the speed by only 5 to 10 percent.
4. If errors start, drop one step and regain accuracy. Accuracy first, speed second.
5. When the loop is stable, widen it a bar at a time to join the flow of the song.
6. Finally, play along with the separated backing at full tempo, or with the whole original.

Splitting practice into 20 to 30 minute blocks tends to keep focus better than one long session; this is general advice and varies by person. If you feel pain or numbness in your wrists, fingers or neck, stop and rest, and consult a professional if it recurs. This is not medical advice. Keep headphone volume low enough that you could still hear someone speaking.

Practice tempo planner

The interactive tool in the Korean section above (Korean UI) builds a table of steps. Enter the original BPM, starting speed, step size, loop length in bars, beats per bar and repetitions per step. For a 120 BPM song starting at 60 percent and rising by 10 percent, it lists 72, 84, 96, 108 and 120 BPM. Move to the next step only after finishing a step without errors.

Playing along and ear training

Backing tracks for ensemble feel

A backing track with your part muted is a virtual band. Unlike a metronome it keeps the song's dynamics and section changes, so you learn from the song itself when to come in and when to rest. First aim to get through the whole song once at a slow speed, then raise the tempo. Practice carrying on after a mistake, because in a real ensemble you cannot stop.

Ear training with isolated parts

1. Solo one part (say the bass) and loop one bar.
2. Before touching the instrument, sing the notes or write them down.
3. Play them and compare with the recording.
4. When you find the notes, return to the original and confirm the same bar really has them; artifacts may have misled you.
5. Note the date and speed of each practiced passage so progress shows.

Chord progressions, fill placement and harmony intervals also become clearer when heard one part at a time. Language learners can use the same idea: solo the vocal, repeat the pronunciation, then add the backing to match intonation.

Recording yourself

What you hear while playing differs from what comes out. Late in a practice cycle, record yourself and compare with the original.

• Record while listening to the backing on headphones. Playing it on speakers puts the backing into the recording.
• A phone voice recorder is enough; record from the same position each time for a fair comparison.
• Listen against the original part and compare one of timing (ahead or behind), pitch, tone or dynamics. Trying to judge everything at once fixes nothing.
• Record the same passage a week apart so progress is objective, and put the date and speed in the file name.
• Wireless earbuds with latency can blur your sense of timing; use wired or low-latency ones if possible.

With Charcoal Player

If you want separation, pitch-preserving speed change and loops in one app, one option is Charcoal Player, a Mac app I build and run. This is the only section about it; the routines above work with any tool. Only features confirmed in the app documentation and store description are listed.

• Source separation into four basic parts (vocals, drums, bass, other instruments); a six-part mode also separates guitar and piano. Those two are experimental and vary by song.
• Pitch-preserving speed from 0.5x to 3.0x, sample-accurate A-B repeat, and a skip interval adjustable from 0.5 to 30 seconds.
• The separation model is downloaded once after consent; after that, separation runs on the Mac and music files are not uploaded. Requires macOS 15 or later and is Mac only.
• Playback, speed change, A-B repeat and separation are free. Exporting separated results is free up to 3 times per 24 hours, then requires an optional subscription (or lifetime purchase). See the store for prices.

The screenshots in the Korean section show the library and transport controls, the six-lane separation panel (a lane reading "not detected" when a part is absent), an A-B loop at 0.75x, the export dialog offering MP3, AAC, Apple Lossless, WAV, AIFF and FLAC, and a Voice mode that splits a recorded conversation by speaker, which is more useful for language or meeting listening than for instruments.

The separation feature assumes you use music you own or have the right to use, for personal practice. Posting, distributing or reusing separated audio in other works may need the copyright holder's permission. Rules vary by country, so ask the rights holder or a professional about your situation. This is not legal advice. Details are in the Charcoal Player guide.

Sources

Same URLs as the Korean list above: Rouard, Massa and Défossez, "Hybrid Transformers for Music Source Separation" (arXiv 2211.08553) [1]; the Demucs project README (github.com/facebookresearch/demucs) [2]. Not verified, so left out: quality comparison figures for individual commercial apps and any quantitative link between practice time and skill, which are described only as general practice advice. The percentages and times in the routines are suggestions and vary by person.

FAQ