NIBBLE — Melody / Percussion

Turns the Workshop System's mysterious four buttons into tools to make melodic lines and lo-fi drum loops.

Four buttons. Fifteen voltages. Ten notes.

A program card for the Music Thing Modular Workshop System Computer that turns the system's Four Voltages module into a playable instrument.

Four Voltages is billed as a "minimum viable keyboard": four non-latching buttons, one knob, four outputs from a resistor network. Press combinations and you get different voltages — but they are arbitrary voltages, they change every time you touch the knob, and nothing downstream knows what to do with them.

NIBBLE learns them. Patch one Four Voltages output into CV In 1, teach the card which voltage each combination makes, and the ten combinations become ten notes of a scale — or a lo-fi drum kit with a looper behind it.

Four buttons is four bits is one nibble. Hence the name.

Watch it work

A two-part demo and walkthrough — calibration, the hold-and-tap technique, and both modes played:

• Part 1 — https://youtu.be/te2GKfyd_0w
• Part 2 — https://youtu.be/mQxKP0NdALI

Two cards in one

The bracketed names are what the modes are called in the source, so the two are easy to match up.

What the LEDs tell you at power-on

The six LEDs are arranged in a 2×3 block, and the top four mirror the Four Voltages buttons throughout the whole card:

0 1        LEDs 0-3 = buttons A B / C D
2 3
4 5

For about a second after power-on, the card says which mode it came up in:

The splash then clears and calibration starts on its own. On the bench it reads as one movement: light up, go dark, begin asking for buttons.

Playing it: hold one, tap the others

This is the technique the card is built around, and it is worth learning first.

Hold C. Tap A. You get a closed hat. Let go of A — silence, no retrigger. Tap B. A snare. Let go — silence again. Tap A. Closed hat again.

The held finger is a bank-select; the tapping finger plays. It repeats indefinitely with no spurious hits, and it works exactly the same in MELODY.

That falls out of how the card solves its central problem. Four Voltages does not return to a rest voltage when you let go — the output sits at the last-pressed button's level. So releasing AB leaves the CV sitting at A's voltage, which looks identical to genuinely pressing A. A naive design fires a spurious note every time you lift a finger.

NIBBLE's rule: when a pair is released onto one of its own two buttons, that one level is ignored. Anything else plays. So AB → release → A is silent, but AB → release → A → B plays B, because the ghost was A, not B. And once the voltage moves anywhere else, the ghost is forgotten — so you can come back to A and it sounds.

Calibrating

The Four Voltages knob changes every voltage it produces, so the card cannot guess and does not try to remember. Teach it whenever you move that knob.

Set it up like this

Use output 1 of the Four Voltages section, with its knob at about twelve o'clock. That gives ten well-spread, reliably separable voltages.

The knob position genuinely matters. It moves all four outputs at once, and towards either extreme several combinations collapse onto nearly the same voltage — at which point no amount of software can tell them apart.

Well-spread voltages are the whole game. Everything else about this card works or fails on that one thing. When two learned levels sit close together the symptoms are not a clean failure but an intermittent one: a combination that plays the wrong note, fires twice, or does nothing at all — and does it differently each time, because a few millivolts of drift is enough to move the reading across the boundary between them. It reads as a flaky card rather than a mis-set knob, which is what makes it worth saying twice.

If you get that, the fix is not to play more carefully. Move the Four Voltages knob and calibrate again, and watch for the collision warning (LEDs 4 and 5 blinking) which tells you exactly when two captures were too close. A calibration that completes with no warnings will behave predictably; one that warns will not.

Doing it

Calibration starts on its own at power-up, since the learned levels are never saved and it is the first thing you would do anyway. You can start another at any time by holding the momentary switch for 2 seconds, and abort with the same hold.

Then, for each of ten steps:

1. The LEDs show which buttons to press. LEDs 0–3 are laid out exactly like the A B / C D buttons, so the pattern is the instruction — no counting.
2. Hold that combination on Four Voltages.
3. Tap the momentary switch to capture it.
4. All six LEDs flash: got it. Next step.

The order is designed for your hand, not for the computer: the four singles, then the pairs as top row, bottom row, left column, right column, and the two diagonals.

The two warnings use different lights on purpose: the button LEDs mean it was your hand, the marker LEDs mean it was the card.

The collision warning is deliberately quiet. It fires only when two levels land within about 95 mV of each other — well below the point where the card actually starts confusing them — so if you see it, it is worth acting on. A warning that appears on combinations you can play perfectly well just teaches you to ignore it.

How it ends, and what to do about it

If it fails, the ten captures spanned less than about 1.2 V — which in practice means nothing is patched into CV In 1, or the cable is dead. The card keeps whatever calibration it had rather than installing a useless one, so nothing is lost by a failed attempt.

If two combinations collide, move the Four Voltages knob a little and run it again. If one output stays stubborn, try one of the other three — they respond differently to the same buttons, and it costs nothing but moving a cable.

Until you have calibrated, the card runs an evenly-spaced guess so it still does something. In MELODY, LED 5 solid means it is running a real calibration; off means it is still guessing.

Learning is not saved across power cycles, deliberately. A stored calibration would be a wrong calibration the moment the knob moved, and a card that looks calibrated but plays the wrong notes is worse than one that admits it knows nothing.

MELODY

Called KEYS in the source. Boots by default.

The ten combinations play ten degrees of the selected scale — not ten pitches. Combination → degree is fixed forever, so re-learning after moving the knob gives you the same notes back. The four bare buttons are the bottom four degrees; adding a second finger climbs.

Panel

In / out

The two audio outputs carry different envelopes — that asymmetry is the point, and it is what Knob X balances. They are for driving an external filter and VCA: patch them at the Humpback filters and an amplifier and one macro knob shapes the whole voice. They run from about 25 ms to over five seconds, so the macro covers clicks through to long swells. Note the audio outputs are DC-coupled and carry CV happily, but they are not calibrated the way the CV outputs are: fine for modulation, not accurate enough for pitch.

The X knob morphs between three relationships: DARK (the filter closes well before the note fades — plucks and muted things), BALANCED (both together), and BRIGHT (the filter outlasts the amplitude, so the note blooms rather than shutting down — pads and swells).

CV Out 2 is a second voice, not a modulation source: the same note a third higher, picked from whichever scale the Y knob is on, so it is major or minor according to where you are in the scale. Patch it to a second oscillator with Pulse Out 1 as a shared gate and the ten buttons play two-part harmony.

The scales

Twelve, on the Y knob, ordered dark → bright from fully anticlockwise to fully clockwise — so the knob reads as a single mood axis rather than a list, and the LED bar (more light = brighter) matches it.

The right-hand column is worth noticing: because the ten combinations are ten degrees, a four-note arpeggio spreads them over more than two octaves while the chromatic scale packs them into three quarters of one. Choosing a scale chooses the range as well as the flavour, and the difference is obvious the moment you play it — the arpeggios are the ones to reach for when you want the buttons to cover real ground.

The root of the scale is 0 V, so an oscillator sitting at its own zero is already in tune with the card — no winding it down an octave to find the keyboard. CV In 2 transposes upward from there, up to two octaves.

PERCUSSION

Called DRUMS in the source. Hold the momentary switch at power-on.

Same calibration, same ghost rule — but the buttons work differently, and this is the important part:

A single button is a SHIFT, not a sound. Only pairs make a noise.

Hold one button down and tap the others. Hold C and tap A, B or D for three different sounds; hold D and tap A, B or C for three more. Keep tapping the same one and it repeats cleanly, as fast as you like.

That is there because percussion is mostly repeated hits on the same drum, and a keyboard reading of the buttons cannot do it: to play AC twice you have to pass back through C, and if C were itself a sound then every repeat would be interrupted by a spurious one. Silent singles make the four buttons into bank-selects you can hold for as long as you want.

Twelve voices, because A-held-tap-B is not the same as B-held-tap-A.

Those two close the same switches and produce an identical voltage — press order simply is not in the signal. What the card can see is the level the voltage came from, which is the button that was already down. Latching that recovers the ordering and doubles the kit for free.

Hold C is the workhorse: with a right thumb on C, A / B / D fall under the fingers as closed hat, snare and kick — a whole beat without moving your hand. Hold D is the tom row for fills. A and B carry the colour, including the crash.

The sounds

All synthesised: a pitched body, a noise component, and an exponential decay on each. Nothing is sampled, which is why the Y knob can reshape the whole kit at once — every number below is a live variable, not a recording.

Pitch is where the body starts and, where it sweeps, where it lands. Fall is how long that sweep takes — the exponential drop that makes the toms read as Simmons-style rather than as short bass notes. Length is how long the voice stays audible. Level is its weight in the mix, and the numbers look lopsided for a reason: they are set by perceived loudness, not by amplitude.

Two effects drive them. The ear is roughly 24 dB more sensitive at 1 kHz than at 55 Hz, so a bright voice needs far less level than a kick to sit alongside it — which is why the cowbell is 16% and the kick 100%. And the ear integrates over time, so a long voice needs less again: the crash is the longest thing here and sits at 9%.

A rough rule if you edit the table in drums.cpp: an octave higher wants about half the level, and twice as long wants about two thirds.

The three cymbal voices — hi-hat metallic, open hi-hat and crash — are ring-modulated against a second oscillator at a deliberately non-integer ratio, which is what makes them clang rather than hiss. The plain closed hat is not, so the two hats sit differently in a pattern.

Y shifts the whole kit together: anticlockwise is lower and longer, clockwise higher and shorter, roughly half to double pitch and ±3 decay steps. It applies when a voice is struck, so sweeping the knob never warps a sound that is already ringing — and its movements are recorded into the loop, alongside the filter. Two automation lanes, so a pattern can change its own kit character as it goes.

Sixteen voices can ring at once, and when they are all busy the card steals the quietest rather than the oldest — so a voice near the end of its decay goes before one that has just started.

Panel

In / out

The click runs whenever the loop does, so there is something to record along to. Patch it at a click voice, or just watch LED 4.

Both audio outputs carry the same mono bus. They are not a stereo pair and there is no setting that splits them: patch either socket and you get the whole kit, which is one less thing to get wrong mid-performance.

Riding a recorded knob

Main and Y are both recorded into the loop, and both work the same way live: moving the knob mutes that lane's playback while you move it and for a quarter-second after. Let go and the recorded sweep picks up again from wherever it has reached.

Nothing is destroyed by touching a knob — grab the filter, ride it through a section, let go, and the pattern carries on exactly as recorded. It is a performance override, not a mode.

While recording, a knob writes only where you actually move it. Sweep across half a bar and that half is replaced; the rest keeps whatever was there. A knob sitting still records nothing, so it can never flatten an earlier sweep just by being armed.

Replacement uses a short window rather than an exact instant — about an eighth of a beat — because no two passes over a knob land on precisely the same moments. Anything within the window counts as "the same place" and gets replaced, so re-recording a sweep genuinely overwrites the old one instead of interleaving with it.

A bassline, for free

The four single buttons make no drum sound, so their CV output was going spare. They play four bass notes on CV Out 1 instead — root, a tone below, the fifth, and the octave — gated on Pulse Out 1 alongside the hits.

CV Out 2 gates it — a 5 V blip on each shift press, so the bass voice can have its own envelope without opening on every drum hit the way Pulse Out 1 does. Live presses only; a recorded pattern has no shift presses in it.

No combos, no scale, no quantiser. Once a pattern is looping, patch CV Out 1 at an oscillator and put a simple line under it with the same four buttons. It also falls out of the shift gesture for nothing: holding a button to reach the kit holds its bass note too, so the root sustains while you play drums over it.

A bonus rather than the feature — but it costs one table and an output nothing else was using.

The looper

Four bars. It records events, not audio, which is what makes overdub lossless — the twentieth pass sounds exactly like the first, because nothing is ever re-recorded. It is also why the tempo knob re-times a pattern instead of pitching it, and why filter-knob moves can be recorded alongside the hits.

The loop stores which sound — kick, snare, crash — not which buttons made it. How a hit was played belongs to the performance, not the pattern, and it means re-arranging the gesture map can never silently change what an old loop plays back.

Hits are quantised to 1/16 on playback; filter sweeps deliberately are not. Punching in with the switch does not reset the loop position.

Patch a clock into Pulse In 1 and the loop follows it — one pulse per beat. It overrides the X knob for as long as pulses keep arriving, and hands the tempo back about three seconds after they stop, snapping to wherever the knob is pointing by then. It locks anywhere from 30 to 240 BPM, and nudges into phase rather than snapping, so locking on never stutters the pattern.

To erase, hold the switch for two seconds to re-enter calibration. That is the erase gesture: every other control is spoken for while playing, and the four singles in particular get held for long stretches as shifts, so anything button-based would fire constantly. Recalibrating is also the one moment a stale loop is guaranteed to be meaningless — the levels are about to change.

Under the hood

• Everything is fixed-point integer. No floats anywhere in the audio path.
• Level detection settles for ~12 ms before it believes a voltage, with a Schmitt band so two close combinations cannot flicker between each other. Triples and the all-four combination are recognised as not being any learned level, so they do nothing rather than guessing at a neighbour.
• Control logic runs at 3 kHz inside the 48 kHz audio callback; voices, envelopes and the filter run at full rate.
• Single core. There is nothing here that needs a second one.
• ~4.6% of flash and ~8.3% of RAM.

Building

Needs the Pico SDK 2.2.0.

cmake -B build -G Ninja
cmake --build build          # -> build/nibble.uf2

Hold BOOTSEL on the Computer, plug in USB, drag the .uf2 across.

There are Python models of the trickiest logic in tools/ — the ghost rule and learn round-trip, the filter's stability, and the looper's event ordering. Run them after touching the corresponding C++.

Status

Working, and played on hardware. Ten rounds of hardware feedback have settled the behaviour: calibration, the ghost rule and triggering were right on first contact, and everything since has been playability — envelope ranges, kit balance, the looper's overdub, the external clock, and how the knobs hand over between your hand and a recorded sweep.

docs/DEVLOG.md records the design decisions and the bugs found along the way, several of which were only visible because the trickiest logic is modelled in Python before it is written in C++.

The idea is portable

The calibration and keypress decoding here are not specific to a drum machine or a keyboard. Any card can learn a set of button-combination voltages and act on them — the ghost rule, the shift-and-tap gesture and the ten-step learn are a general technique for making Four Voltages playable. Expect siblings.

If you are building one, docs/LESSONS.md is written for you: what to lift wholesale, the bugs that recur on this platform, and the things that only a hardware session teaches.

Credits

ComputerCard by Chris Johnson (Music Thing Modular), MIT.

The drum voices are generalised from Wild Pebble (Workshop System release 74); the scale tables are from Lockstep (release 89); the degree-indexing approach is from CA Sequencer (release 19); the Schmitt-band quantiser idiom is from Goldfish (release 11); and the state-variable filter's update ordering is from bends (release 45). Thanks to all of them.

Built with Claude Code.

Licence

CC-BY-4.0 — use it, fork it, sell it, just credit Andy Jenkinson (uglifruit). ComputerCard.h is Chris Johnson's and keeps its own MIT licence.