Play the kick, then the snare, on the drums below. They sound like they belong together: the snare sits a few notes above the kick, both are dry, both stop at about the same moment. Now drag Snare tune all the way up and play them again. Each drum is still a perfectly good sound, but the pair no longer sounds like one drummer's kit. It sounds like two records playing at once.
Play the kick and the snare, then move Snare tune far from where it starts. At what point does the snare stop belonging to the kick?
Two drums, one kit
Then try Kit tone. It changes every drum at once, darker or brighter, and the two keep belonging to each other. A drum kit is more than eight good sounds. It's a handful of decisions the sounds share.
You'll describe a kit and generate candidate hits with a model, then choose and prepare the ones that hang together. Map them in a sampler so they respond to how hard you play, and finish with a beat and a saved kit for your next track. Before you generate anything: if you picked the single best-sounding kick, snare and hat from fifty candidates, each judged on its own, would you get the best kit?
Musician steps, model names
Stage 1: Describe the kit before the drums
Every kit you love was made for a kind of music. A dusty boom-bap kit sounds like it was sampled from old records in a small room. A techno kit is tight, dry and punchy enough to cut through a club system. A lo-fi kit is soft at the edges. The sound comes from the job.
So write the job first, in one sentence: the tempo, the space and the feeling. "A dry, warm kit for a slow hip-hop beat at 85 BPM, like it was recorded in a small wooden room." That sentence is your kit's identity. Every hit you keep later should fit inside it.
Then write one line per drum, starting with the shared sentence and adding only what's different about that drum:
- Kick: short and round, low but not boomy, a soft click.
- Snare: a tight crack with a short body, no long ring.
- Closed hat: a small, crisp tick.
- Open hat: the same hat, open, ringing for about half a beat.
- One percussion sound: a rimshot, a shaker or a wooden knock to fill gaps.
Repeating the shared words in every line matters. A model only knows what you tell it each time, and the shared words are what pull the hits toward one room and one era. In machine-learning terms, your description is the conditioning: the input that steers what the model generates. Here, the conditioning is your words.
Checkpoint: one sentence for the kit and five lines for its drums. Drum programming explains what job each drum does in a groove, if you're not sure what the snare needs to do.
Stage 2: Generate more options than you need
Now ask a model for sounds. Generate several takes of each line, five or six snares rather than one. Each take starts from a different random starting point, called a seed, so the same words come back as different snares: one brighter, one with more body, one with a strange little rattle. That spread is what you're after. You can't choose a family of sounds if you only have one of each.
In Maru, you can do this while building a sampler bank in Studio: in Sources, describe a sound or the whole kit, generate takes, preview them and choose what to add. Generation needs to be set up in your Studio first; the sampler-bank guide walks through it. Any audio generator that accepts a text description will do the same job. Name every file by its role and take ("snare-03") as it arrives, so you can still find the good one an hour later.
There's a second way to generate, and it's worth trying. Instead of asking for one hit at a time, ask for a short recording of the whole kit being played, hit by hit with gaps between them, then cut it into single hits. In Studio, the agent can split an attached recording into separate one-shots at the silences.
Our take: hits that come out of one generated take often already share a room, because the model imagined them in the same space and through the same recording chain. Hits generated one by one can each come from a different imaginary studio. Neither way is always better, so try both and listen.
If a hit sounds wrong for reasons the words can't fix, don't spend twenty minutes rewording the prompt. Generate more, or change the sentence that describes the kit. Audio generation explains what a prompt and a seed really control.
Checkpoint: five or more candidates for each drum, named by role and take.
Stage 3: Choose a family, not five winners
This is the stage that makes the kit. Put a simple beat on a loop (kick on 1 and 3, snare on 2 and 4, closed hats on every eighth) and swap candidates into it one slot at a time. Never choose a drum in solo. A snare that sounds huge alone may cover the kick in the beat; a modest one may sit perfectly.
Listen for four shared decisions:
- Pitch. The kick and snare have a pitch, even if you can't name it. When the snare sits comfortably above the kick, the pair sounds intentional. You heard this in the first lab.
- Decay. How long each hit rings. Drums in one kit belong to one family of lengths: in a tight kit, nothing rings much past the next hit.
- Tone and room. Brightness, grit and the space around each hit. One drum recorded in a hall makes the others sound like they're in a cupboard.
- Loudness. The balance between drums. The hats should sit under the snare, not beside it.
One snare rings too long
When one hit sticks out, it usually breaks exactly one of the four. Name which one before you replace it. "Too much room" means generate a drier one; "too bright" might mean a gentle filter is enough. Now the kit sounds like one drummer in one room, not a compilation.
Same idea, different place: when your ear decides a snare belongs with a kick, it's doing the job the field calls similarity, which models do by comparing short lists of numbers. Tagging and similarity shows how, and why a model's idea of "alike" isn't always the sound that does the job.
Checkpoint: one kick, one snare, a closed and an open hat and one percussion sound that work in the loop, plus a second-choice take for each drum. Keep the second choices; Stage 5 uses them.
Stage 4: Trim the starts and keep the balance
Generated hits rarely start exactly where the sound starts. There's often a sliver of silence or a soft lead-in before the attack, and a few milliseconds is enough to make a drum feel late against the beat. Trim each hit to just before its attack, and let it fade out over its last few milliseconds so nothing clicks when it stops. Preparing samples shows where to cut and why a cut a little too late loses the punch.
Two things are particular to a kit. First, don't normalize the hits: turning every file up to the same peak makes the closed hat as loud as the kick and throws away the balance you chose in Stage 3. Set levels by ear in the loop, with the kick and snare leading and the hats tucked under them. Second, listen for a generated tail that carries its own reverb into the next hit; shortening it is often easier than finding a drier take.
Checkpoint: every hit starts on its attack, ends without a click and sits at a level that works in the loop.
Stage 5: Map it so it plays like a drummer
Load your hits into a sampler. Put them on the standard drum notes so any drum pattern or MIDI file plays them correctly: kick on note 36 (called C1 or C2, depending on your software), snare on 38, closed hat on 42, open hat on 46, and your percussion sound on any free key nearby. In Studio, a sampler bank does this in Mapping; the Maru Sampler below plays a bank the same way, from pads or keys.
Play one pad softly, then hard, then four times fast. Does it respond like a drum, or like a button?
Play a kit from pads
Maru Sampler
Mapped instruments and drum kits in your browser.
- Level 0.80
- Release 300 ms
Now make it respond. Two settings turn a set of files into an instrument:
- Velocity layers. Put the second-choice snare on soft velocities if it's gentler, and the main one on hard velocities. Playing softly then sounds different, not only quieter. If you don't have a softer take, generate one: "the same snare, played softly".
- Round robins. Give each drum two or three takes that alternate on repeated hits, so a fast hat pattern doesn't sound like a machine gun. Takes that came from the same prompt are ideal: same drum, slightly different hit.
Last, put the closed and open hat in the same choke group, so the closed hat cuts the open one short, the way a drummer's foot closes the hi-hat. Now play a fast hat line with a few accents. It should feel like a hand, not a stuck key.
Checkpoint: every drum on its note, soft and hard hits that differ in tone, and alternating takes on repeated hits. Key and velocity mapping and Round robins and articulations go deeper on each.
Stage 6: Play a beat on it
Program or play eight bars: the plain loop from Stage 3, then a variation with a ghost snare, an open hat at the end of a phrase and a fill in the last bar. Play it at the tempo you wrote in your sentence from Stage 1.
This is the real test, and it's often where one more problem appears. A percussion hit that sat fine in the plain loop may clash once the fill arrives, or the kick's tail may blur the bass line you add. If something bothers you, go back one stage, not three: swap in a second choice, shorten a tail, rebalance a level.
If the kick never quite works, you don't have to generate one. A synthesized kick from Kicker or the drum synth at the top of this page can sit in a generated kit, as long as it shares the kit's decisions: similar length, similar weight, tuned below the snare. Mixing sources is fine. Mixing rooms is what breaks a kit.
Checkpoint: an eight-bar beat that sounds like one kit from the first bar to the fill.
Stage 7: Save the kit and its story
Save the bank and give it a name that says what it's for: "Small Room 85" tells future you more than "kit-final-2". In Studio, Save bank keeps a revision you can return to, and Export bank gives you a file you can reopen later.
Keep a short note with it: the sentence from Stage 1, which tool and settings generated each hit, and the seeds of the takes you kept. That's the kit's recipe. It lets you make a matching second kit later ("the same room, but brighter") and it tells you, and anyone you share it with, where every sound came from. Before you publish or share the kit, check the generator's terms for what you're allowed to do with its output. Rights and maintenance covers what to check.
Checkpoint: a saved kit with its recipe, and the eight-bar beat played on it. Bring it to your next track and it's ready, and it already sounds like one instrument.
Quick check
- You pick the best kick and the best snare from twenty candidates each, judged in solo. In the beat, the snare seems to float above the kick in a different space. What happened? Each was chosen alone, so nothing made them share a room. Choose in the loop, and if the snare has more reverb than the kick, swap it for a drier take rather than adding reverb to everything.
- Your generated snare is programmed exactly on beat 2, but it always sounds a touch late. The tempo is right. What would you check? The start of the file. Generated hits often begin with a sliver of silence or a soft lead-in before the attack. Trim it to just before the attack and the snare lands on the beat.
- A fast hi-hat pattern sounds mechanical even though every hit sounds good. What do you change? Add round robins: two or three takes of the same hat that alternate, and give soft and hard hits different takes. The repetition of one identical recording is what your ear catches.
Next steps
- Go deeper: Generating notes. You let a model suggest the sounds. A different kind of model can suggest the pattern, and its output is notes you can edit.
- Use it: Beatbox to drum pattern. Play your new kit with your mouth: beatbox a groove and turn it into a part for these drums.
- Sideways: Drum synthesis. Build the same kit from oscillators and noise instead of a model, and hear which decisions carry over.













