Drumkit notes

From Helpful
Revision as of 20:54, 16 July 2020 by Helpful (Talk | contribs)

Jump to: navigation, search

DIY electronic drumkit

The idea: build a cheap electronic, velocity-sensitive, MIDI-output drumkit.

There are many simple and cheap DIY ideas out there. Check the video sites.

I found this a nice project, in that it has analog and digital and physical things to figure out, educational frustrations along the way, and satisfying results.


You may care about a few different things

getting enough energy to the sensor, and consistent amount of energy to it
more makes it easier to reliably measure impact and its velocity
detecting velocity
and getting it reproducably
getting minimal energy to neigbouring pads - i.e. physical isolation from others
avoid false triggers (without having to hard-ignore pads, which would be bad for real drumming)
allows us to be more sensitive
allows some choice between finger drumming and stick drumming (there's a large energy difference)
allowing fast drumming, without false triggers
we'll discuss why this is a thing later

having it be useful for practice
pads that aren't loud
half the reason is fewer complaining neighbours
relevant for stick drumming, not finger drumming
getting the pads to physically bounce roughly like their real-world counterpart
do hihat pedal properly
and potentially things like muting

Physical design

A few of those can be solved largely in physical design, making your life easier once you get to the electronics and code.

Also, if you only want finger drumming, then life is much easier because the forces are much smaller, and so barely carry to neighbouring ones. I'd recommend your first version be this.

If you want to play as hard as regular drums this is a little harder to design well.

Say, you probably do still want thing to stay in place, but a very solid construction will also be better at hard-coupling energy to the other sensors. This is okayish since drumming is mostly fairly hard, but if you want higher sensitivity this is a design tradeoff you have to deal with.

If you don't care about the physical loudness of hitting pads, or the physical bounce, then a piezo on the back of a piece of wood already works fine.

If you do care about the bounce and loudness, and about crosstalk and the false triggers that may cause on closeby pads, then the materials above and below matter.

Good physical isolation is slightly harder than you'ld think, if you want an otherwise quite solid thing. (For example, I at one point had a MIDI guitar, which has piezos detecting the picking of each stringalike, which if you pick them hard enough will trigger the next one over even though they're mounted in rubber. Could be lessened with filtering by velocity at MIDI level, but that a commercial product has this issue is an indication of the problem) Note that this also dampens the resonance of the pad itself, so the signal strength (and shape), so imply some recalibration in the velocity detection part.

Many designs use a piezo sensor because they're cheap and work pretty well. For finger drumming the constuction can be small, for real drumming you probably want decent-sized pads for aiming reasons, and the implied large surface is nice for a consistent amount of energy and resonance to measure, which also makes it a little easier to get a more predictable amount of energy.

Keep in mind that on a solid board, the main resonant node is typically in the middle, so that's where you get best response.

Also keep in mind that you want to attach it in a way that doesn't fall apart immediately. A dollop of wood glue is likely to fail sooner rather than later. Something entirely solid holding it in might damage the piezo over time. Something soft inbetween may dampen the response a little, for better or worse.

Also keep in mind that the wires on the piezo also carry shock directly to the piezo, so you may want to guide those away with a bit of foam - depending on your design.


Dealing with the piezo's capacitance

Piezos are essentially stress sensors, so respond well to weight, and to impact and the vibration that causes.

Due to their design, they can be seen as stress-to-voltage sensors in parallel with a capacitor (nanoFarad scale(verify)) (and a very large resistor, GOhm scale(verify))).

That combination means it will hold the stress-generated voltage for a relatively long time, to the point that if you hook up a piezo directly to and ADC, you get what looks like a varying resting voltage between hits. Resting a weight on the pad will give a proportional high value. (You may even get small changes over time and with temperature as the piezo and/or the thing it's on expands or shifts)

You can write code to deal with such shifty input, but it's messier, and there's more edge cases where you can get wild triggering.

So what many designs do instead is add a resistor across the piezo.

This is essentially a discharge resistor for the capacitive effect, which means the voltage output is now primarily the changes in stress, and will be near 0V the rest of the time.

You can quantify the impulse response, because it resembles a lowpass filter(verify). Since a piezo disc's capacitance will probably be in the ballpark of 500pF to few dozen nF, and you probably want dropoff within a few dozen milliseconds, something on the order of 100kΩ makes sense.

More widely something in the range of 10k to 1M can make sense, depending on further component and design choices. (In a more portable design you may wish to add a ~1M trimpot, in series with ~10k as a minimum, so you can easily calibrate this later)

If you are multiplexing a single ADC (example being an arduino, but it's likely given cost of having ~10 distinct ADCs) you are likely to find the signal is cross-bleeding between the channels, giving responses on channels you're not hitting.

This largely because ADCs like the Arduino's are designed to deal best sensing things with lowish impedance. (<10kOhm for AVRs, see section "Analog Input Circuitry" in the datasheet). If the impedance of what you're measuring is more than that, then the ADC's sample-and-hold capacitor starts interacting enough to be part of the sampling.

One thing you can do is lower the discharge resistor. The problem with that is that even with a ~10kOhm discharge resistor, which is a lot better, you still haven't solved it, and it comes at the cost of a weak and faster-dampening signal, making it harder to be sensitive to velocity, and sensitive in general.

A better fix is presenting a low impedance to the ADC via a buffer (which also implies the discharge resistor now only affects the dropoff), such as with a FET or an op amp.

My personal preference is the op amp because the tweaking the gain is a little easier (and it's slightly easier to breadboard, with two or three quad op amps for 8 to 12 channels). That said, op amps come with further considerations, like that the output voltage swing is usually lower than the rail, so try to keep under that effective clipping level.

(Other fixes are dedicated ADCs, or even MCUs per channel, but that makes communication a little harder.)


  • Note also that you can only multiplex so much before you get few oscillations per channel and start aliasing and velocity starts being less robust. You may want to run the ADC faster, as the time resolution is more important than the noise.

Protecting your input

Diodes you see in a bunch of circuits seem to typically be ~5V zeners from ground to clip peaks, protecting the ADC/buffer from larger piezo voltages - a few dozen volts if hard-coupled and hit hard.

How necessary this is depends on the choice of piezo and resistor, and specs of ADC/buffer, but it's a good idea in that it can help lifetime, and costs a few cents.

Detecting empty sockets


If you don't care about velocity, then you can probably write the code in a few minues.

"Do analogRead, Is it above threshold and have we not triggered this in the last 100ms? Then trigger MIDI note" Goes a long way.

When you want velocity, and smartness to help isolation, then things become a little more interesting.

Isolation and sensitivity

If the physical part of isolation wasn't perfect, you'll still see a little of the force from neigbouring pads.

If nearby hits arrive at much lower intensity, then a little threshold tweaking will help a lot.

Still, the better that the physical isolation is, the lower that threshold can be, so the more we can support softer playing even while you're hitting something nearby harder.

You also have the option of 'if two pads seem to hit less than a few ms apart, and one with rather lower intensity, then then maybe just trigger the strongest'.

This will sometimes ignore actual playing, and punish good timing, so ideally you improve the isolation instead.


Velocity makes things more interesting.

There are various methods. Some are simple, some are complex, some are slightly faster (latencywise), some are more consistent, some are handier for softer playing, etc.

For example, we could watch channels for the initial hit. Once we decide it has been triggered, we keep sampling it as long as the values keep increasing, trigger once it starts falling, with the peak value as the velocity

This is probably the lowest-latency you can get, and it's simple code in that it doesn't need to store much. However, high frequency content makes this sometimes stop too early/low, and using just the first peak may be a little less robust between similar hits depending e.g. on the resonance of your design. If/when (just) that first peak clips, it will always report the same value

Another way is to, once we decide to trigger, keep sampling for another 2-15 milliseconds, and average the values (or e.g. over-threshold time), and use that as the velocity. This fairly robust to any clipping of the initial peak.

Because tiny peaks that were only just enough to trigger the "could this be a hit?" threshold, but the average would be very low, it's a little easier to reject it as probably-crosstalk or noise, while still allowing soft hits.

has a fundamental tradeoff between longer (more consistent velocity) and shorter sampling time (lower latency)
can be done without storing samples in RAM (keep adding magnitudes to a large-typed counter, and divide by sampling time)

Apparently there's some that try to find the envelope of the signal. Probably more accurate, but slower, and harder to do

Since you don't have direct control over what pad voltage should correspond to the softest hardest hits (depends on the size of the piezo and the size of the pad), you may want to calibrate.

A 'play a few sample hits on everything, and store that into EEPROM' is a simple and effective idea.
This also makes it more a lot more modular/portable around varying physical designs.

That also makes it easier to change the curve away from linear by applying a function (e.g. a power).

Fast playing, hard playing, and false re-triggers

Playing harder means the signal is stronger, but also a slightly longer time it takes to taper off the oscillations, up to 40ms or so.

If you're reading fast enough, then after a note trigger you may notice the not-quite-gone peaks from the same hit and may trigger again.

A conceptually simple fix is to add the condition "AND did I not trigger in the last 100ms?" in your note-triggering code. {{{1}}}

However, things like flams and drumrolls are hits on the order of milliseconds apart, so suddenly you need to detect new hits before the last one has fallen off, so e.g. tell whether a half-waves was falloff from the last hit or a new one.

This is harder to get right, and while possible involves some tradeoffs.

One decent approach is to normally leave a low threshold for triggers from nothing, but after a note trigger temporarily set a higher threshold, base it on the last emitted's strength.

One issue is determining that value reliably - you probably want it lower than the first peak, but higher than the average.

This means you can leave the general threshold low for lower-latency triggers, evade most dampened peaks from the same hit, but still catch most similar-velocity hits.

Consider also that for some channels, you can reject such things harder. You're not going to do few-ms drumrolls on a kickdrum, on a crash it's more of a sustain so the speed doesn't matter as much, while on a ride drum or tom this is more important.

If you vary designs, e.g. have kickdrum just be a block of wood, then you may want some knobs to tweak thresholds and curves per channel -- though you probably have dedicated channels anyway (because otherwise you need to think about configurable MIDI mapping).

If you do velocities, then you may like the ability to tweak the curve on the fly.

Hihat logic

If you care to practice for real hihats, you want to include the pedal.

It should emit

open-hihat when sensor is hit while open, note number 46 (name: A#1)
pedal-hihat when pedal closes, note number 44 (name: G#1)
closed hihat (42 in GM) when the sensor is hit while closed, note number 42 (name: F#1)

(those notes are from the GM percussion set, it might vary somewhat. And remember the octave numbering is not well set within midi, it may be +- 1(verify))

Since the pedal should allow staying closed, a basic footswitch works well enough. In theory you could combine it with a piezo to also get velocity.

Two-sensor logic

Muting logic

MIDI practicalities

(see also MIDI_notes#Percussion)

In my code I delayed sending NoteOff for ~10ms (unless the next trigger is sooner), mostly for debug reasons - it makes it easier for to see hits when something (physical/DAW) chooses to visualise it.

MIDI Velocity

You may find that compared to the sound you get at velocity 100ish, velocities under 30 or so are barely heard - even ghost notes would probably be at 40-60.

Much lower and it'd sound like you're playing part of your kit from the next room.

(I spent a day trying to figure out why hits went missing when it was just that velocity ~15 is not audible)

As such, you've got some amount of balancing between enough variation expression while also getting it to be repeatable, sound natural, and and allow ghost notes.

This, may take good isolation, calibration, habit, and it may help to put a transform that makes the mapping from physical to MIDI velocities nonlinear.

A little tweakability to that can't hurt.

PC side

Drum sounds

This article/section is a stub — probably a pile of half-sorted notes, is not well-checked so may have incorrect bits. (Feel free to ignore, fix, or tell me)

If you want to get decent sounds for free, try

that VST has velocity senitivity and sounds pretty good, and is free though has a click-once-at-startup annoyance (that goes away when you contribute)
e.g. Cantabile[1] is free, but more on the performance than DAW side. There are other options.
tell it to use the WASAPI output, 48kHz (or higher) if you can, and you can usually get away with 256-sample buffer on generic modern hardware, for ~6ms latency which is pretty decent for drumming.

Linux has a steeper learning curve when it comes to MIDI routing, and low latency audio output. It's on my list of things to figure out -- later.