Diamond Stem Studio

AI stems · true instrumentals · a cappellas · karaoke. Everything runs in your browser.

🎵

Drop an MP3 here or click to choose (WAV, M4A, FLAC, OGG also work)

Tip: name files “Artist - Title.mp3” so Karaoke can find lyrics automatically.

Separation engine

Model: not loaded GPU: checking…

The AI engine runs Meta’s HTDemucs on your own machine. It’s fastest in Chrome or Edge with WebGPU; without a GPU it still works, just slower (a 4-minute song can take 10+ minutes).

Cleanup (bleed control, applies to every engine)

Bleed suppression Vocal ⇄ Instrumental focus

Load a song to begin.

Diamond Stem Studio: the manual

This app takes a finished, mixed song and pulls it back apart into four stems: Vocals, Drums, Bass and Other (guitars, keys, strings, synths, pads). It also builds an Instrumental (Drums + Bass + Other) and an A cappella (Vocals only). Separation always happens in two stages:

Separation engine: does the heavy lifting of deciding which sound belongs to which stem.
Cleanup: a fast finishing pass that removes leftover bleed. It can be re-run in seconds without separating again.

Quick start

Drop your MP3 in the Studio tab.
Pick AI · Demucs (the first run downloads the model once).
Click Separate stems, then use the Instrumental / A cappella buttons to listen.
Hear bleed? Move the Cleanup levers and click Re-apply cleanup.
Export the stems, or go to Karaoke to sing.

The three engines

Engine	How it works	Use it when
AI · Demucs	Meta’s HTDemucs neural network, trained on thousands of real multitrack recordings. It recognizes what a voice sounds like, not just where it sits in the stereo field.	Always, when you want the cleanest result. This is the one that fixes “I can still hear the vocals”.
Spectral	Finds sound that is identical in the left and right speakers (lead vocals are almost always mixed dead-center) inside the vocal frequency band. Drums are found as short, sharp events, bass as low sustained notes.	Quick previews, slow computers, or when the model can’t download.
Classic center-cancel	Subtracts the entire center of the mix inside the vocal band, the same trick as the “karaoke” button on 1990s stereos.	Old or simple stereo recordings where the vocal is dry and centered. Expect centered instruments (snare, bass guitar mids) to get thinner too.

Every lever, explained

Spectral levers (only used by the Spectral and Classic engines; the AI makes these decisions itself):

Lever	What it does	Vocals	Drums	Bass	Other
Vocal isolation strength	How perfectly centered a sound must be to count as “vocal”. Higher means stricter.	Higher: cleaner vocal but thinner. Lower: fuller vocal with more instruments leaking in.	—	—	Lower strength pulls more centered keys/guitars out of Other.
Vocal low cut	Nothing below this frequency is treated as vocal.	Raise it (200–250 Hz) to keep kick and bass out of the vocal. Lower it for deep male voices.	Protects kick drum.	Protects bass.	—
Vocal high cut	Nothing above this is treated as vocal.	Lower it (7–9 kHz) if cymbals leak into the vocal. Raise it to keep “s” and breath sounds.	Protects hi-hats and cymbals.	—	—
Bass crossover	Tonal sound below this goes to Bass, above it to Other.	—	—	Raise it for bass with lots of growl or upper harmonics. Lower it if piano or guitar low notes end up in Bass.	Mirror image of Bass.
Drum sensitivity	How eagerly short, sharp sounds are classed as drums.	—	Higher catches ghost notes and hats; too high pulls guitar strums and piano attacks in.	—	Lower keeps pick and hammer attacks in Other.
Quality	High uses 4× overlapping analysis windows, Fast uses 2×.	High = smoother, less “watery” sound. Fast = about twice as quick.

Cleanup levers (work on every engine; click Re-apply cleanup after changing them):

Lever	What it does	Guidance
Bleed suppression	Compares all four stems at every frequency, many times a second, and gives each slice of sound more decisively to whichever stem is loudest there. 0 = off, 2 = standard (Wiener filter), 4 = maximum.	Raise it when you hear ghosts of one stem in another. Too high (3.5–4) can sound “swirly” or underwater on sustained notes. Sweet spot: 2–3.
Vocal ⇄ Instrumental focus	Tips close calls one way or the other. Positive = any doubtful sound goes to Vocals, so the Instrumental gets cleaner. Negative = doubtful sound goes to the band, so the A cappella gets cleaner.	For a true instrumental, use +2 to +4. For a true a cappella, use −2 to −4. Neutral 0 is best for full four-stem exports.

Mixer: M mutes a stem, S solos it (only soloed stems play), and the slider sets its volume. Click any waveform to jump there. The mixer volume affects listening only. Exports are always full level.

Recipe: a true instrumental (no vocal ghosts)

Engine: AI · Demucs.
Bleed suppression 2.5–3, Vocal ⇄ Instrumental focus +3.
Separate, then press Instrumental and listen on headphones through the loudest chorus. Choruses with stacked harmonies leak the most.
Still hear whispers? Focus to +4 and re-apply cleanup. If the band starts sounding hollow, go back one step.
Export Instrumental.wav.

Why a few traces can survive: reverb and delay on the voice are smeared across time and frequency and look a lot like pads, and backing vocals are sometimes treated as “Other”. The +focus setting targets exactly these.

Recipe: a true a cappella (vocals only)

Engine: AI · Demucs.
Bleed suppression 3, Vocal ⇄ Instrumental focus −3.
Separate, press A cappella, and check quiet verses for hi-hat or snare ticks.
Hear drum ticks? Suppression to 3.5. Voice sounds robotic? Focus back to −2.
Export A cappella.wav. Tip: you can run the song twice (once at +3 for the instrumental, once at −3 for the vocal) and export each, because the two goals pull in opposite directions.

Recipe: a karaoke track

Make a true instrumental (above).
Go to Karaoke, search the lyrics, and press Sing.
Beginners can raise Guide vocal level to 10–20% to hear the melody underneath.
Karaoke mix in the Studio export keeps a faint guide vocal (−20 dB) baked in.

Karaoke mode

Search lyrics: looks up the song in LRCLIB, a free community lyrics database that stores time-stamped lyrics. Results with synced ✓ scroll automatically. The one whose length matches your file is listed first.
Paste: plain lines or LRC format ([01:23.45] line).
Tap-sync: for plain lyrics. Play the song and tap Enter as each line begins. When you’re done the lyrics are synced and can be saved as a .lrc file for next time. The app also remembers them for this song.
Timing offset: if the highlight runs early or late, slide it. Negative shows lyrics sooner.
Guide vocal level: 0% = pure instrumental. Raise it to hear the original singer quietly.
The dots at the top count you in before the first line and after long instrumental breaks.

Live remix & echo

After separating, the Live Remix Recorder turns the mixer into a performance instrument. Press ● Record remix (or R). The song restarts from the playhead and everything you do is recorded: mutes, solos, volume moves, presets and echo. The take ends when you press stop, pause, or the song ends.

Control	What it does	Performance tip
`1`–`4` / M	Mute or unmute Vocals, Drums, Bass or Other.	Drop the drums (`2`) for a breakdown, then bring them back on the downbeat.
`Shift+1–4` / S	Solo: only soloed stems play.	Solo the vocal for an a cappella intro, then un-solo to slam the band in.
FX → Echo On	Turns on the delay for that track.	A touch of 1/8-dotted echo on vocals is instant dub.
Amount	How much of the track is sent into the echo.	15–35% for space, 60%+ for an obvious effect.
Time	Gap between repeats. Synced values (1/16 to 1 bar) follow the Tempo; Free lets you set milliseconds. Changing it while repeats ring gives a tape-style pitch swoop.	1/4 for rock and soul, 1/8 dotted for reggae and dub, 1/16 for a fluttery slap.
Feedback	How many times it repeats (0% = one echo, 90% = nearly endless).	Keep under 60% normally. Push it up for build-ups; the limiter protects you.
Tone	Darkens each repeat (low-pass filter).	2–4 kHz sounds like vintage tape echo. 10 kHz+ is crisp and digital.
Hold to THROW / `A S D F`	While held, the track is sent 100% into the echo with extra feedback.	The classic dub move: hold `A` on the last word of a line, then mute the vocal. The echo tail keeps ringing after the mute.
Tempo	Auto-detected from the drum stem after separation. Type to correct it.	If echoes feel off-beat, try doubling or halving the number.
Master volume / Limiter	Overall level. The limiter stops echo build-ups from clipping.	Leave the limiter on for masters.

Downloads per take: Studio master WAV re-renders your performance offline from the exact list of moves you recorded, so it’s click-free and sample-accurate even if your computer stuttered while you performed. Live capture WAV is a direct recording of what came out of the speakers; use it when you jumped around the timeline during a take. Record as many takes as you like and keep the best.

Troubleshooting

Problem	Fix
Vocals still audible when muted	Switch to the AI engine. Then raise Focus to +3 and Bleed suppression to 3.
Watery, swirly, “underwater” sound	Lower Bleed suppression (to 1.5–2). On Spectral, set Quality to High.
AI model won’t download	Check your internet connection, or download the model file from the link in the status message and use Load .onnx from disk. After one successful load it’s cached in your browser.
AI is very slow	Use Chrome or Edge (WebGPU). The GPU pill shows whether it’s available. Without a GPU, expect roughly 2–4× the song’s length.
Lyrics not found	Try just the title, check the spelling, or paste them and use Tap-sync.
Mono file warning	Spectral and Classic need stereo. The AI engine works on mono too.

What to expect

No separator in the world is perfect, because a finished mix destroys some information. The AI engine with cleanup gets very close to studio stems on most modern songs. The only source of a guaranteed 100% clean instrumental is the original multitrack session or an officially released instrumental.