Sunday, July 22, 2018

Using the midi tuning specification (MTS) standard in supercollider


I have a hardware synthesizer Dave Smith Instruments Rev2 ( (aka DSI Rev2)with support for MTS (=midi tuning specification) and I'm intrigued by its possibilities. How can I reprogram my synth's frequencies from supercollider? In a next post I may or may not (depending on how fast I can solve some annoying bugs :) ) explain how to download scala tunings and keyboard mappings with the Rev2 from supercollider, but for now let's concentrate on the low-level MTS messages. Most of the information in the article explain calculations, so it should also be usable for anyone trying to use MTS from any programming language.


It will be explained how to assign random frequencies to all midi notes by composing a valid sysex message containing bulk tuning information. The explanation is somewhat math heavy, but every step is completely detailed which should help you implement something in your favorite programming language.


Midi tuning specification is a protocol that allows users of synthesizers to assign any frequency to any midi note. You can reverse your keyboard if you like, or you can assign random frequencies to all keys, or you can explore the wonderful world of microtonality and xenharmonics. The way to assign any frequency to any midi note is by sending midi sysex messages.


Sysex is a hook in the midi protocol that allows for sending anything over midi. "Anything" is to be interpreted as follows: "anything that fullfills the demands made on bytes sent over MIDI, i.e. the highest bit is always 0". This constraint means that in some cases data has to be encoded in a form that is suitable to be sent over midi, and this is also the case for MTS messages. We'll go over the calculations in excruciating detail in the rest of the article.

With some exceptions, sysex messages are not standardized. Any manufacturer can implement its own sysex messages. DSI, e.g. uses sysex to perform OS updates on their instruments. Such sysex messages typically are not publicly documented.

The sysex messages used for sending tuning information, on the other  hand, are standardized. They come in a few flavors. There's a message to request an synth to send over its tuning to a computer, but the DSI instruments appear not to respond to those messages, so I guess they are not supported.  There's another message to send tuning information from the computer to a synth, and that works well. Still other messages exist to retune a few notes in real-time, but I haven't tried those yet.

MTS message structure

MTS message structure is not very complicated, but it takes a bit of explanation. The full explanation can be read in the midi tuning standard spec, but it might contain some mistakes (currently under investigation).

First of all let's explain the general message layout. 

F0 F7 7E 7F 08 01 IDX C1 C2 .. C16 TA1_0 TA2_0 TA3_0 .. TA1_127 TA2_127 TA3_127 CS F7
  • F0 F7 appears at the start of any sysex message
  • 7E specifies that it's a non-real-time message.
  • 08 specifies that we're about to send tuning information
  • 01 specifies that we're about to send a bulk request, i.e. tuning information for all midi notes
  • IDX specifies a 0-based tuning index, this corresponds to the tuning index selected in DSI's global parameters menu, alternative tunings menu entry
  • C1 .. C16 are 7-bit ascii codes for the characters in the name of the custom tuning
  • TA1_0 TA2_0 TA3_0...TA1_127 TA2_127 TA3_127 are the actual tuning information. Tuning information for each midi note is described using three bytes, to be explained below.
  • CS is a checksum, also explained below
  • F7 signals the end of the sysex message

TA bytes tuning information structure

How do the three bytes TA1, TA2, TA3 encode tuning information for a midi note? Well, the first of the three bytes, TA1 is a number that says which lower midi note number is closest to the desired frequency, if the instrument would be tuned in 12-Tone equal temperament (i.e. freq  =  2(midinote-69)/12*440 Hz or vice versa midinote  =  12*log2(freq/440 Hz)) + 69, where midinote is a number 0-127, and freq is a float.

The bytes TA2 and TA3 then encode the difference in "cents" between the frequency indicated by the midi note number in TA1 and the desired frequency. The difference in cents between two frequencies f1 and f2 can be calculated as nc  =  1200*log2(f2/f1)), but supercollider provides some built-in functions to hide these formulas.

In the MTS message, 128 such triples TA1,TA2,TA3 are sent, one for each midi note. 

So if you wanted to retune the 50th midi note to the value of 670 Hz, you'd have to put as 50th triplet in the message the following bytes:

  • the closest, lower midi number corresponding to 670 would be 76 (midi note 76, on a 12-TET tuned instrument, corresponds to a frequency of  2(76-69)/12*440 = 659.26 Hz ). The way to calculate this 76 is by taking the floor of the formula that transforms frequency to midi note number: floor(12*log2(670/440 Hz) + 69) = 76. In hexadecimal, 76 is 4C. 
TA1_50 = 76 (or hex: 0x4C)
  • the difference in cents between 659.26Hz and 670Hz would then be 1200*log2(670/659.26) = 27.98908618457 cents. This difference in cents will always be < 100, because the reference frequency is the closest lower midi note (and the difference between two successive midi notes in 12 TET, is always exactly 100 cents).
  • We have 2 bytes TA2 and TA3 to encode this number 27.98908618457. Because in midi bytes, the high bit is reserved to indicate commands instead of data, that gives us a total of 14 bits (= 0 to 2^14-1) to encode a number in the interval [0, 100] (I include 100 because the difference in cents might e.g. be 99.995). In other words, you get a resolution of 100 cents / (2^14) = 0.0061 cents. That is what we call super high precision :) 
  • If you map number 27.98908618457 from the interval [0,100] to the interval [0, (2^14-1)], by means of the formula 27.98908618457 * (2^14-1)/100 ,you get 4585.4519896181, or after rounding 4585. This 4585 is the number that must be encoded in TA2 and TA3. This happens with bit masking operations:
TA2_50 = 4585 >> 7 = 35 (or hex: 0x23)
TA3_50 = 4585 & 127 = 105 (or hex: 0x69)

Checksum calculation

  • The checksum is calculated as a XOR between all the bytes in the message, except for the sysex header (F0 F7), the checksum itself (obviously) and the end of sysex byte at the end (7F).
  • Gotcha: at the end of the checksum calculation, be sure to "and" the result with 127 (=0x7F) to mask out the highest bit, otherwise you end up with an invalid MIDI message. This is not explicitly mentioned in the MTS standard, but it's implicitly assumed for all MIDI data bytes.


  • Beware of limited precision of floating point operations. In some cases, it may lead to unexpected results and invalid sysex messages as a result. I encountered such problems in my supercollider implementation of the tuning message and put in a workaround.
  • Be sure to check and double check that midi note numbers used as TA1 do not become negative or >127. Failing to do so will cause you hours of debugging (I should know by now!) to find out why sometimes - seemingly out of the blue - the synthesizer does unexpected and weird things when sending a new tuning (like randomly switching to a different program and changing some LFO settings) or why sending a new tuning doesn't seem to have any effect.
  • The MIDI MTS spec states that sending a tuning value with TA1, TA2, TA3 = 0x7F 0x7F 0x7F should be interpreted by the synth as "do not change tuning", but it seems that DSI instruments interpret it as "tune to the highest possible frequency" instead. For this reason the code currently just puts the standard note where no note should be present.

Supercollider code

Note that the following screenshot is part of a supercollider class, part of a larger system to communicate with the DSI prophet rev2 hardware synthesizer (the code can be found here: ), so you'd have to change some minor syntax details to get it working as a regular function (click the image to enlarge, or - even better - copy paste it from github (which may have a more recent version):

Friday, May 11, 2018

Panola: a supercollider PAttern NOtation LAnguage


In supercollider, one of the best ways to schedule notes over time is by using the pattern system. The pattern system is very flexible but the flexibility can make it a bit hard to use. 

In a typical pattern specification, also known as a Pbind, every dimension to your sound event is independent from every other event. This is just a fancy way of saying that you need to specify volume separately from duration, separately from midinote, etc.

If you want to delete a note in the middle of your piece, you have to dig through all these keys in the pattern specification to find the right stuff to remove or add so as to make sure that all the rest still sounds as intended.


By grouping all musical information together and extracting the required keys from that central specification, we can change information in one place and keep all the resulting keys in sync automatically.

Thinking about this, I thought back of a creation I made last year: the midi specification language (MISPEL, see ). I dreamed about having this system available in supercollider. I didn't quite get there yet, but now at least a subset of Mispel is available in supercollider and that subset is known as Panola (pattern notation language), available from 


(For now this is a literal copy of the tutorial on - feel free to go there, it has syntax coloring :) )

// Panola is a way to extract Pbind keys from a concise specification.
// This makes it easier to compose "traditional" music with Pbind, with a lot less
// headache trying to keep the different keys in sync
// It's the type of system I've missed since my day one with supercollider.

// First things first. To install Panola:


// Now you can get the help document by typing ctrl+D with the cursor on the word
// Panola in the next line"a4");

// Let's start with the "Hello world" of Panola: a simple scale.
// The numbers indicate octaves.
// You don't need to repeat octave numbers if they don't change between notes.
~ex ="c4 d e f g a b c5");
~player =;

// asPbind takes a synth name as parameter (which defaults to \default).
// So the above is equivalent to
~ex ="c4 d e f g a b c5");
~player = ~ex.asPbind(\default).play;

// instead of calling a single "asPbind" you can also extract all information separately
// like this you have optimal flexibility in what you want to use from Panola
~ex ="c4 d e f g a b c5");
~pat = Pbind(\instrument, \default, \midinote, ~ex.midinotePattern, \dur, ~ex.durationPattern, \amp, ~ex.volumePattern, \tempo, ~ex.tempoPattern, \lag, ~ex.lagPattern, \legato, ~ex.pdurPattern);
~player =;

// You can make chords using angular brackets. Only note properties of the first
// note in the chord (other than octave number and note modifier (see later)) are
// taken into account.
~ex ="<c4 e> <e g> <c e g c5>");
~player =;

// You can use modifiers on the notes:
// # for sharp, x for double sharp, - for moll, -- for double mol
~ex ="c4 d- e f# gx a# b-- c5");
~player =;

// With underscores you can indicate rhythm.
// The last used rhythm value is reused until a new one is specified:
// Here's four quarter notes (_4) followed by four eighth notes (_8).
~ex ="c4_4 d e f g_8 a b c5");
~player =;

// You can use one or more dots to extend the length of the rhythm, as in traditional notation.
~ex ="c4_4. d_8 e_4 f g_16 a_4.. b_4 c5");
~player =;

// You can also use multipliers and/or dividers to change the length.
// E.g. here we use it to create a note that lasts for three eighths
// (c4_8*3) and to create tuplets (e_8*2/3 f g). Remember that last
// duration/rhythm indication is reused until a new one is specified.
~ex ="c4_8*3 d_8 e_8*2/3 f g f_16 e f e g_4 b_4 c5");
~player =;

// Now we come to the animated property system. We can attach properties to the notes and animate them over time.
// For now two types of animation are supported: linear interpolation and fixed value.
// To indicate linear interpolation, use curly brackets {}. E.g. here we let the tempo gradually increase from 80 bpm to 160 bpm:
~ex ="c4\\tempo{80} d e f g a b c5\\tempo{160}");
~player =;

// Different properties can be combined. Here we let the volume go up until the middle of the phrase, then let it go down again,
// while tempo is rising from 80 bpm to 160 bpm.

~ex ="c4\\tempo{80}\\vol{0.2} d e f g\\vol{0.9} a b c5\\tempo{160}\\vol{0.2}");
~player =;

// If you want to use the fixed values, use square brackets instead. You can switch between fixed and animated everytime
// you specify a new property value. In the next example, tempo remains at 80 bpm until we come to note a. At that point,
// it jumps to value 100 bpm and gradually increases to 200.
~ex ="c4\\tempo[80] d e f g a\\tempo{100} b c5 d e f g a b c6\\tempo{200}");
~player =;

// Using pdur (think: played duration), we can indicate the difference between staccato and legato.
// Here we slowly evolve from very staccato to very legato:
~ex ="c4_8\\pdur{0.1} d e f g a b c5 d e f g a b c6\\pdur{1}");
~player =;

// Using lag we can modulate lag. This can be a way of creating a rubato feeling.
// Linear interpolation is not ideal for this purpose, but it's better than nothing at the moment.

~ex ="a5_8\\tempo[120]\\lag{0} b c6 a5 e d c5 d e c a4 g#4\\lag{0.5} "
"a4_8 b c5 a4 e d c4 d e c a3 g#3 a b c4 d e g# a_2\\lag{0}");
~player =;

// In addition to using predefined properties like tempo and lag, you can also use user
// defined properties, e.g. here we animate a property called "myprop".
~phrase ="c d\\myprop{0.1} e f g a\\myprop{0.6}");
~pattern = ~phrase.customPropertyPattern("myprop"); // extract only myprop values as a pattern
~stream = ~pattern.asStream;{
| i |;
// make a pbind in which the myprop appears as one of the keys, with a default value of 0 for myprop
~pbind = ~phrase.asPbind(\default);
~stream = ~pbind.patternpairs[13].asStream;{
| i |;
// make a pbind in which the myprop appears as one of the keys, with a customized default value of 0.4 for myprop
// (such default values are used if no values for myprop are specified yet, e.g. in the beginning of a Panola string,
//  before any myprop is defined).
~pbind = ~phrase.asPbind(\default, custom_property_defaults:Dictionary.newFrom(["myprop", 0.4]));
~stream = ~pbind.patternpairs[13].asStream;{
| i |;
// make pbind in which only the standard panola keys are included
~pbind = ~phrase.asPbind(\default, include_custom_properties:false);

// These custom properties can be e.g. used to drive synth arguments
// The 303 synth used below is reused from
// which in turn is based on code from Lance J. Putnam
var line;

SynthDef (\sc303 , {  arg  out=0, freq=440, wave=0, ctf=100, res=0.2,
sus=0, dec=1.0, env=1000, gate=1, vol=0.1;
var  filEnv, volEnv, waves;
volEnv =  EnvGen .ar( Env .new([10e-10, 1, 1, 10e-10], [0.01, sus, dec],  'exp' ), gate, doneAction:2);
filEnv =  EnvGen .ar( Env .new([10e-10, 1, 10e-10], [0.01, dec],  'exp' ), gate);
waves = [ Saw .ar(freq, volEnv),  Pulse .ar(freq, 0.5, volEnv)];
Out .ar(out,  RLPF .ar(  Select .ar(wave, waves), ctf + (filEnv * env), res).dup * vol);


line =
"a2_16\\wave[0]\\vol{0.05}\\tempo{120}\\res{0.2}\\sus{0}\\env{1000}\\ctf{100} a a a1 a2 a a3 a2 a a a1 a2 a3 a2 b- g\\res{0.05}"
"a2_16\\wave[0] a a a1 a2 a a3\\sus{0.2} a2 a\\ctf{3000} a a1 a2 a3 a2 b- g\\res{0.2}"
"a2_16\\wave[0] a a a1 a2 a a3 a2 a a a1 a2 a3 a2 b- g\\res{0.01}\\sus{0}\\env{10000}\\ctf{10}"
~player = line.asPbind(\sc303).play;

Saturday, January 13, 2018

Baking sound in supercollider


I want to synthesize gorgeous pad sounds in supercollider. This is possible using additive and/or subtractive synthesis, but it takes a lot of CPU power.


't Is the season to be bakin' fa-la-la-la-laaaaa la-la-la-laaaaa. How about we pre-render the sound into a wavetable and then loop the generated wave table? This moves most of the work to startup time. In the blender 3d program, pre-rendering heavy calculations is known as "baking".

But... won't we lose real-tome creative possibilities then? After all, you cannot modify the partials in a prerendered wavetable anymore?

Well... yes, you will lose some creative possibilities, but we can easily reintroduce some by adding amplitude envelopes, filters and filter envelopes and effects like reverb or phasing and flanging.


One of the more exciting pad generation algorithms no doubt is Paul Nasca Octavian's PadSynth algorithm. It is explained in detail here. I will use it to demonstrate the approach.

In essence it takes as input a basic sound, specified as a set of partials with associated relative volumes. 

It then enriches that spectrum by replacing each of the partials with a fuzzy region containing smeared out partials. This is the equivalent to adding many slightly detuned copies of the original signal, which is how sounds can start to sound a lot warmer. Since human hearing is less sensitive to small frequency differences for high frequencies than for low frequencies, the spreading of partials is made wider in the higher frequency bands than in the lower frequency bands. PadSynth also sets the phases of each of the detuned copies of the signal to random values which helps in reducing undesired comb filtering effects.

PadSynth in supercollider

Supercollider does not come with an implementation of PadSynth, but it is not too difficult to convert the pseudo-code from the detailed explanation into working code. A post by Donald Craig on the supercollider mailing list provides an excellent starting point.

I modified the algorithm a bit as follows: I calculate two different wavetables per octave over 12 octaves (this is probably a bit overkill - feel free to change it). This is an intensive calculation and takes quite a lot of memory. You can certainly reduce this calculation time and required space by reducing how many octaves you want to use, by reducing the number buffers per octave and by reducing the size of the pregenerated wave table (I generate wavetables of around 5 seconds long each).

While playing, the pitch can be chosen by looping over the generated wave tables with a different speed. By varying the speed on-the-fly one can also implement pitch bending or glissando's. All of this is illustrated in the sample code below, in which I will be using supercollider's excellent pattern system to drive a cello-like PadSynth pad. The code can also be found on

Driving the synth from supercollider patterns

Because the synth loops over one of many pre-generated buffers, we'll have to inform the synth about which buffer to use for a given note. For this reason you'll see some calculations being done in the Pbind and Pmono in the sample code. If you decide to generate less buffers per octave, you will also have to adjust these calculations.

More in detail: the parameter \mynote contains a midi note number. This is the note we will want to hear. From that parameter, another parameter \myreference is calculated. This contains another midi note, namely the closest midi note number that was used as a reference to pre-render a wave table: I generate 2 wavetables per octave and each of these wavetables have a reference midi note number. This information is needed in the synth to determine how fast to loop over the buffer. \referencefreq contains the same information but expressed as a frequency in Hz instead of as a midi note number. Remember that I generated two wavetables per octave, for 8 octaves. The parameter \buffer contains the buffer number to use. The buffer being used is the one that was generated with a reference midi note closest to the desired midi note. This information is needed in the synth to know which buffer to loop over. 

The full code is a bit long to paste into this article, so please head over to to get it!
If is down, there's still github as a backup :)