Rodolfo Cangiotti

I'm currently planning to make a Max project that should be able to follow a real musician playing an instrument with a kind of improvised accompaniment. In order to get this result, I thought to focus my project on the utilization of the 1st order Markov Chains through the "anal" and "prob" objects. The project should work with MIDI data only and the parameters that I would like to make a state transition matrix are pitch - semitones and octave - intensity and duration. The most difficult parameter to analyse and to organize into a matrix seems to be the last one - duration: to lighten the amount of the possible states in the matrix (I mean that also an impercetible difference of some ms is taken by the "anal" object as a different state) , I have organized a little subpatch that is able to translate whatever duration of a note in ms to only 30 different states that reminds the classical notation - e.g. crochet, quaver, semiquaver related to a defined time signature. Despite this reduction seems to work pretty good, I am actually facing another problem: when more than a single note are played at the same time - in other words, polyphony - this system is not able to measure the real effective duration of each note and to store it into the "prob" matrix because, before the actual end of a first note, the instance of a new note makes a new time counting to start. I thought to create parallel counters using the "poly" object but, supposing that the first voice can have a longer duration than the second, third, etc. this could alterate the real succession of duration states to be taken from the "anal" object. Can you suggest me a solution to this issue, that is to allow this temporal measurement withouth interfere with the right succession of states and viceversa?

Waiting for your suggestion, I thanks in advance.

Hello all,
I'm currently planning to make a Max project that should be able to follow a real musician playing an instrument with a kind of improvised accompaniment. In order to get this result, I thought to focus my project on the utilization of the 1st order Markov Chains through the "anal" and "prob" objects. The project should work with MIDI data only and the parameters that I would like to make a state transition matrix are pitch - semitones and octave - intensity and duration. The most difficult parameter to analyse and to organize into a matrix seems to be the last one - duration: to lighten the amount of the possible states in the matrix (I mean that also an impercetible difference of some ms is taken by the "anal" object as a different state) , I have organized a little subpatch that is able to translate whatever duration of a note in ms to only 30 different states that reminds the classical notation - e.g. crochet, quaver, semiquaver related to a defined time signature. Despite this reduction seems to work pretty good, I am actually facing another problem: when more than a single note are played at the same time - in other words, polyphony - this system is not able to measure the real effective duration of each note and to store it into the "prob" matrix because, before the actual end of a first note, the instance of a new note makes a new time counting to start. I thought to create parallel counters using the "poly" object but, supposing that the first voice can have a longer duration than the second, third, etc. this could alterate the real succession of duration states to be taken from the "anal" object. Can you suggest me a solution to this issue, that is to allow this temporal measurement withouth interfere with the right succession of states and viceversa?
Waiting for your suggestion, I thanks in advance.


markov-chain-for-live-performance

you probably described the project correctly, but mosst people here are too lazy to read texts, we are more used to patches.

maybe you can post a first attempt patcher of yours, oso that one can see in what direction it will go.

i hope you are aware that when you want to include pitch and duration, that with a second order chain you will already have 4 states because the shit multiplies.

your duration problem first of all originates from the fact thaat you dont know when a note will end when it starts. if i was you, i would give up the idea of building the state/transition table in realtime. what if the first note is 2 minutes long? first play everything in a buffer, then convert the delta times into duration and then analyse the markov stuff in nonrealtime.

your duration problem first of all originates from the fact thaat you dont know when a note will end when it starts. if i was you, i would give up the idea of building the state/transition table in realtime. what if the first note is 2 minutes long? first play everything in a buffer, then convert the delta times into duration and  then analyse the markov stuff in nonrealtime.

Well, thanks for suggestion, Roman. Attached to this post you can find the two .maxpat file I am working on.

Anyway, I have two questions related to your additional thoughts: first, what do you mean with the fact that I should have four states already defined if I want to use a second order Markov chain? And then, what do you mean with Delta Times?

EDIT: with Delta Time do you mean the temporal interval between two Note-On messages? Because I am reading the help file of the "borax" object just now and it is defined in this way.

Well, thanks for suggestion, Roman. Attached to this post you can find the two .maxpat file I am working on.
Anyway, I have two questions related to your additional thoughts: first, what do you mean with the fact that I should have four states already defined if I want to use a second order Markov chain? And then, what do you mean with Delta Times?
I wait for your reply.

EDIT: with Delta Time do you mean the temporal interval between two Note-On messages? Because I am reading the help file of the "borax" object just now and it is defined in this way.


i meant by this that you are going to create a lot of data if you want to include the durations in the states collection. there might be C4 notes of 15 different durations in a piece and each will end up as one state.

"second order" means that every state in the table consist of two states: the next note chosen is not only based on the last note, but on the last two notes. so the last state will be a list like "60 65"

"delta time" is the relative distance to time 0. this is, for example, how midi works. a note start at 25,000 seconds and ends at 27,000 seconds. it does not have any "duration", it just has a begin and an end.

what you said about "between note ons" is also true. this might be useful, too, for this application. it would be a bit like converting the input to a mono-legato version and then measure the duration.

"delta time" is the relative distance to time 0. this is, for example, how midi works. a note start at 25,000 seconds and ends at 27,000 seconds. it does not have any "duration", it just has a begin and an end.
what you said about "between note ons" is also true. this might be useful, too, for this application. it would be a bit like converting the input to a mono-legato version and then measure the duration.

Ok, thanks again; I wait for a kind of feedback from your side about the patches.

Well guys, I have worked a bit more on this project and in order to update the thread I attach the last patch: to partially solve the issue above mentioned I have created a kind of table where all the data are provvisory stored and analyzed before to be written into the matrix. Furthermore I have upgraded the order of the chains from 1st to 2nd order. The result seems better but there are still little issues like in-stuck states and more, "irregular" timing, etc. that are actually only partially solved. What about? Can you also suggest me some improvements?

Markov Chain for Live Performance