OMax

T h e S o f t w a re I m p ro v i s e r

© Martin lartigues

Version 3
MacIntel / Max 5

Documentation by G. Assayag & G. Bloch, Oct 2009

Omax is designed and developed by The OMax Brothers: G. Assayag, G. Bloch, M. Chemillier in col-
laboration with Shlomo Dubnov
http://omax.ircam.fr
 TABLE

Introduction: OMax and the Stylistic Reinjection Process!.........................3

System Requirements!.....................................................................................5

Quick Start!.......................................................................................................6

The Three Worlds of OMax: Midi, Event-audio and Spectral-audio!............6

Tutorial 1. Simple Midi Processing!................................................................8

Tutorial 2. Controlling your Improvisation!..................................................10

Tutorial 3. More Improvisation Control: Phrase Segmentation!.................13

Tutorial 4. Oracle Housekeeping!..................................................................14

Tutorial 5. Going Audio!.................................................................................16

Tutorial 6. Sound Output Control!.................................................................19

Interlude 7. Audio Buffers!.............................................................................20

Tutorial 8. Going Spectral!.............................................................................21

Spectral audio, quality and continuity!........................................................22

Tutorial 9. Block Impros and Loops!.............................................................23

Tutorial 10. Presets for Improvisation Control!............................................25

Ref1: Audio Control Panel Reference!..........................................................27

Ref2: MIDI Control Panel Reference!............................................................29

Ref3: Recording Setup Reference!...............................................................30

Ref4: Spectral Control Panel Reference!.....................................................32

Ref5: SuperVP Control Panel Reference!.....................................................35

Ref6: Miscellaneous!......................................................................................37

Bibliography: OMax related Publications!...................................................40

Introduction: OMax and the Stylistic Reinjection Process

Machine improvisation and related style learning problems usually consider building repre-
sentations of time-based media data, such as music, either by explicit coding of rules or ap-
plying machine learning methods. Stylistic learning of musical style use statistical models of
melodies or polyphonies to recreate variants of musical examples. Recent advances in learn-
ing techniques indicate that particular behaviors can be emulated and that credible behavior
could be produced by a computer for a specific domain.

In the field of music improvisation with computers there has been recently notable advances
in statistical music modeling that allows capturing stylistic musical surface rules in a manner
allowing musically meaningful interaction between humans and computers. The contributors
to OMax have experimented with several of these models during the past years, and recently
cristallized the results of their researches in OMax, an environment which benefits both from
the power of OpenMusic (Assayag, Agon), for modeling and high level programming, and
MaxMSP (Puckette, Zicarelli) for performance and real time processing. OMax allows for
interaction with a human player, on-the-fly high-level feature extraction and stylistic learning,
virtual improvisation generation, stylistic model archiving and hybridizing. It operates on two
distinct time scales: the Max one, which is close to real time and involves fast decision/
reaction, and the OpenMusic one, which has a deeper analysis/prediction span over the past
and the future. These two conceptions of time interact and synchronize over communication
channels (OSC), through which musical data as well as control signals circulate in both direc-
tions. A decisive advantage we have found in this hybrid environment experience is its
double-folded expandability. In the OM domain, it is easy, even while the system is running,
to change the body of a lisp function and test incremental changes. Furthermore, it is easy to
enrich the generation by connecting the system to a wide variety of compositional algorithms
available in this environment. The same thing is true in the Max domain, with a comprehen-
sive collection of real-time generation and processing modules. We think of this setup more as
an indefinitely modular and extendible experimental environment for testing new ideas about
interaction, than as a fixed application. However, in the setup documented here, there is only a
subset of OpenMusic involved (we don’t need the graphics), materialized as a small footprint
lisp image, and a Max patch that can be run on top of Max or even on top of Max Runtime if
you don’t own Max.

OMax provides a virtual musical partner that learns all its knowledge from the musicians it’s
playing with, in a non-supervised mode, and that is fitted to a real-time audio context as well
as Midi. One way of looking at the interaction that emerges between the musician and OMax
is to characterize it as a process of Stylistic Reinjection. The musical hypothesis behind stylis-
tic reinjection is that an improvising performer is informed continually by several sources,
some of them involved in a complex feedback loop (see Figure below). The performer listens
to his partners. He also listens to himself while he’s playing, and the instantaneous judgment
he bears upon what he is doing interferes with the initial planning in a way that could change
the plan itself and open up new directions. Sound images of his present performance and of
those by other performers are memorized, thus drifting back in memory from present to the
past. From the long-term memory they also act as inspiring sources of material that would
eventually be recombined to form new improvised patterns. We believe that musical patterns

3
are not stored in memory as literal chains, but rather as compressed models, that may, upon
reactivation develop into similar but not identical sequences: this is one of the major issues
behind the balance of recurrence and innovation that makes an interesting improvisation. The
idea behind stylistic reinjection is to reify, using the computer as an external memory, this
process of reinjecting musical figures from the past in a recombined fashion, providing an al-
ways similar but always innovative reconstruction of the past. To that extent, the virtual part-
ner will look familiar, as well as challenging, to his human partner. The interesting thing is to
see how the human partner reacts to his “clone” and changes his improvisation accordingly.
Top improvisers, who have used this system at an early prototype stage (Bernard Lubat, Mike
Garson, Philippe Leclerc, Hélène Schwarz, Guerino Mazzola and others) have developed their
own way of interacting, as can be seen and listened to on the OMax web site:
http://www.ircam.fr/equipes/repmus/OMax.

Pragmatically speaking, OMaxMax (the Max component) “listens” to a musician partner

through Midi or audio channels. It extracts high-level features from the incoming signal and
segments it into events and phrases. OMaxMax feeds continuously OMaxLisp (the OM/Lisp
component) with a stream of Midi-like information (we call this augmented Midi). On its
side, OMaxLisp continuously builds a memory model of the sequence of events. This Model
is called an Oracle, based on works on string sequence modeling by Crochemore, Allauzen
and al. (Oracle have been used, among others, in order to discover patterns in DNA strings).
OMaxLisp is also ready to continuously “improvise”, that is browse the model in order to
generate variant sequences that it sends as a continuous augmented Midi flow to OMaxMax.

A lot of interface controls let the OMax operator steer the virtual improvisation by navigating
into the memory of the session, from the immediate present to the far past. As a session can be
archived and continued later, this far past could be weeks or years ago. As oracles can be hy-
bridized, the memory model could be trans-individual as well and incorporate pieces of the
Repertoire. Using these controls, the operator will be able to conduct a co-improvised per-
formance with one or several human musicians.

We leave it to the bibliography to document the many technical aspects underlying the OMax
technology.

4
4
OMax has been mainly developed by Gérard Assayag, Georges Bloch and Marc Chemillier.
The stylistic modeling researches underlying it have been carried mainly by Gérard Assayag
and Shlomo Dubnov. The spectral version of OMax owes a lot to Shlomo Dubnov’s pioneer-
ing work. OMax uses Yin~ the pitch tracker technology by Alain de Cheveigné (Max imple-
mentation by Norbert Schnell). The Spectral Mode uses FTM/Gabor library by the Ircam
Real-Time Musical Interaction Team.

System Requirements
OMax 3 runs on Max 5 on Apple Macintosh with Intel processor
Spectral OMax requires the (free) Ircam FTM library.
Pitch/Time transformations need the Ircam SuperVP for max library
OMax is multi-channel: if the audio hardware allows it, one can play back on 8 channels.
Several instances of OMax can be run simultaneously on different inputs using the patch Mul-
tiOMaxMax

5
5
Quick Start
OMax is constituted of two main programs: OmaxLisp and OmaxMax

The OmaxMax program is the same for both platforms, there is a just a single modification
for communicating with Lisp. If you own max/MSP, you can modify the program yourself as
explained later).
Launch OMaxLisp and OMaxMax. OMaxMax opens two windows, one for general com-
mands and one improvisation commands and display. Keep the windows close to each other.

The Three Worlds of OMax: Midi, Event-audio and Spectral-audio

Before any use of OMax you have to decide if you want to use it in Midi mode or in Audio
mode. Furthermore, Audio mode is declined in two ways: note-events or spectral descriptors.
Therefore OMax is divided into three worlds
1. Midi mode: you have a Midi input (e.g. a performer playing a Midi keyboard) and a
Midi output (OMax improvising to a Midi port connected to e.g. an external Midi ex-
pander or a software expander or a sequencer).
2. Event-audio mode: you have one (preferably two, see further) audio input through
Max dac1/dac2, and OMax will be improvising by recombining the audio captured at
the input and sending it to one to eight audio output channels. In addition, you’ll be
also able to listen simultaneously to 2 Midi outputs corresponding to the direct audio
input as analyzed by the pitch tracker, and to the OMax improvisation.
3. Spectral-audio mode: you have one audio input through Max dac1/dac2, and OMax
will be improvising by recombining the audio captured at the input and sending it to
one to eight audio output channels.

How to choose a mode

The modes are chosen on the popup menu of the Initialisation panel.
Midi: choose MIDI_NoAudio and press OK.

6
6
 Event-audio: different presets correspond to internal set-
tings carefully tuned for selected input instruments
(Tenor Sax, Bflat Clarinet, Bass Clarinet, Voice, Percus-
sion, Sound Poetry-spoken voice). The « Default » audio
item is a good average setting to start with. The “piano-
bar” set is designed for the Moog pianobar, or any audio
instrument equipped with its own Midi detector. The Bass
Clarinet also works for double bass.
Spectral-audio: there are two choices, spectral_mfcc (Mel frequency cepstral coefficient) and
lpc (Linear predictive coding). For the moment lpc gives better results. [Remainder: you
NEED to have FTMLib installed in order to be able to use the spectral settings].

When switching from one mode to another, the Red top left panel actually changes: choosing
a mode activates a script that modify the patch organization. Advice is given to do this once at
the beginning of a session, just after loading the patch, an not to fool around switching modes
while the patch is running. You can save-as your patch (inside the folder it came from) and it
will remember its Midi-or-Audio configuration. So, for instance, you could have a patch spe-
cifically for the voice and you won’t have to configure it anymore.
Before proceeding to Tutorial 1, go into Midi Mode (Don’t forget to press OK).

7
7
Tutorial 1. Simple Midi Processing
Setting up OMax for MIDI.
1. As said above, go into MIDI mode by Chosing MIDI_NoAudio in the popup menu of the
Initialisation panel
2. Since, in MIDI, there is no waveform to be displayed, chose “number” in the upper right
popup menu of the improvisation command window. A simple timeline is displayed.

Starting the improvisation.

On the figure below, we have highlighted the 4 zones of interest for launching a simple im-
provisation process.

3. Use the Midi port menus to select your input and your output (warning: even if the correct
ports are displayed, it is usually necessary to reselect them when restarting the program).
4. Press OMaxInitSession. This will initiate the communication between OMaxLisp and
OMaxMax. The OMaxLisp listener and OMaxMax will react as pictured below (the left
window will a little bit differ in an Intel Mac)

8
8
 .
5. Check the toggle box Record Enable.
6. If there’s Midi flowing at the input, OMax will begin (almost) instantly to reimprovise it,
sending it’s improvisation on the selected out port.
7. Use start / stop to control the improvisation, as well as the level slider, and the Mute/0dB/
Cue little buttons nearby the slider.

If you would like to keep OMax improvising, but stop « listening » to the input (and thus
« learning » new material), just toggle Record Enable off.
If you would like to start another session, forgetting all the material learned so far, push the
OMaxInitSession button again. The session will be reinitialized, and the toggle Record Enable
will stay in the state (on or off) it was in.

9
9
Tutorial 2. Controlling your Improvisation

If Tutorial 1 was succesful, you have now Midi streaming in (e.g. a Midifile, or a musician
playing a Midi keyboard) and OMax generating a co-improvisation simultaneously. Now, you
would like to control this co-improvisation. This is the purpose of the Panel above.

We have already seen start and stop, which just starts or stops the OMax Improvisation (but
without preventing OMax from listening and learning).

The Continuity slider helps to control the density of recombination in the OMax improvisa-
tion. If set to 8 (default), it will, in average, replicate 8 successive musical units from the
learned sequence before trying to recombine, that is jumping to another place in the sequence.
It is good policy to set it to a higher value (e.g. 16) when the input is dense (high tempo, lot of
notes per time unit) and even to a higher value for spectral improvisations (30-80 are decent
values in spectral mode). Of course this is just a statistical indicator: OMax has a very sophis-
ticated scheme for looking at the best combinations, so it could choose to increase the conti-
nuity internally in order to favor a better location to jump. Very low values could result in a
chaotic behavior, which may be of interest.

The Quality radio buttons sets the quality of recombination. High quality recombination
(lower values of the button, e.g. 1) jump between places that share a long common musical
context, and that are rhythmically more coherent. One should prefer higher quality, of course,
but, depending on the type of music at the input, high quality could result in no recombination
at all (thus just replicating the input) because OMax doesn’t find jumps that match your exi-
gence. The default value is 3; you should use 2 as much as possible, and 1 is adapted for some
types of music with high redundancy rate. There is a debug mode to inform you about the at-
tained rate of recombination that will be described later.

The Regions radio buttons lets you control in which part of the input sequence learned so far
OMax should peek musical material to recombine. They can be seen as a control of the mem-
ory of the improvisation (long term, short term).
• All means the whole sequence. In this mode, OMax navigates freely in the totality of the
material learned so far.
• Region limits the available material to a certain time region of the sequence learned so far.
Time region are selected using the horizontal rslider (see below) or directly on the wave-
form display (in audio oracles).
• Follow is a particular case of a region that is continuously redefined in order to follow the
real-time input. The length of this region is set by the numbox just right to the follow but-
ton. With a length of 5 seconds, OMax will always be re-improvising the material that has

10
10
 been played during the 5 last seconds, following, as in a pursuit, the performer, while he/
she moves ahead in time.
Note that choosing Follow for a while, then choosing Region will « freeze » the current fol-
low region: OMax will cease to follow the performer and instead navigate in the freezed re-
gion. If the region is small, this is a particularly interesting way of installing a groove derived
from very recent material, on top of which the performer will feel at ease in elaborating a cho-
rus.

There is also an experimental follow mode, that you can try by opening the VFollow abstrac-
tion (bottom-right of the OMaxMax patch) :

You can set a time period (default 5000ms) and a follow percentage (default 40%). When you
check the VFollow mode, OMax will divide the given time period in two successive halves
whose duration depend on the given percentage. During the first half, OMax will be in follow
mode, then try to connect to some material in the All mode for the second half, then back to
follow mode and so on. The purpose of this script is to give the feeling of a close follow proc-
ess while enriching the improvisation with patterns brought back from the past, just as a hu-
man musician could do.

The long horizontal Rslider represents the whole session from the beginning (left) to the cur-
rent time (right). This Rslider can display a waveform when in audio mode. It is a linear map-
ping of the Oracle itself, which you can see as a sequence representing in an ordered way the
music that has been input since the beginning of the session. By selecting regions with the
mouse in this slider, you will select input material that has been played at some point in the
past, and force OMax to re-improvise this material, instead of wandering freely into the whole
sequence.
Because the slider is of fixed size, time is continuously scaled to fit in as the session moves
ahead, so a 20 seconds region could occupy a smaller or greater horizontal space depending
on the overall duration of the session so far. The total length of the oracle is indicated on the
right side of the Rslider.

11
11
Regions selected in the rslider become active only when the Region radio button is pressed.
Two numboxes indicate the current region start and end time in seconds/milliseconds. The
bigger numbox at the far right indicates the duration in seconds of the sequence learned so
far.

Play a while with this interface until you feel comfortable, then go to Tutorial 3.

12
12
Tutorial 3. More Improvisation Control: Phrase Segmentation
As the input from the performer is flowing in, it is analyzed into “phrases”. Phrases are just
musical sequences surrounded by pauses of a certain length (default 1000 ms). This is a more
convenient way than continuous time in seconds to represent things and to remember specific
moments of interest. This segmentation threshold can vary between 500 and 2000 ms and be
modified on the segmentation panel:
If you check the tog-
gle Phrase Limits in
the improvisation
panel, then your se-
lection in the rslider
will be automatically
aligned with phrase
boundaries. The two
green numboxes will
indicate the phrase
range. If you just
click (without drag-
ging) in the rslider,
you will select the
phrase surrounding
the point in time corresponding to the click, and its number will be indicated in the green
numboxes.

For instance if you remember that the performer has played three long phrases since the be-
ginning, it becomes extremely easy to select by a simple click, say, the second phrase, then
force OMax to improvise on that phrase with the Region radio button.

Whether your are in phrase mode or in continuous time mode, you might want to trigger on
and off the improvisation several times (using start and stop) and have OMax begin every
time at the same place before going into some improvised phantasmagoria. This is the mean-
ing of the toggle bordr strt (start on border). Actually OMax will always restart on the first
event of the current selected region. Every time you restart, OMax will begin with the same
phrase, replicating the performer for a while (depending on continuity) then digress into his
own improvisation, doing differently so for each restart.

If a region is selected and you start an improvisation in mode All, the chosen region will be
used to choose a starting point. After the starting point, OMax is then free to improvise wher-
ever it likes. So, even if the All mode favors a completely region-free improvisation, you can
at least decide where to start, especially if you check bordr strt.

13
13
Tutorial 4. Oracle Housekeeping

The musical material input by the performer, taken as a sequence in time, is modeled in
OMaxLisp into a data structure called a Factor Oracle. Basically, it is the sequence itself, plus
a certain number of arrows indicating where and how to jump. Thus, the Oracle is the long-
term memory of the session, including a way to represent the logical organization of patterns.

OMaxLisp can handle several oracles at a time, for instance, the oracle of the current session
plus recorded oracles of previous sessions, and it can switch from one to another. It can even
learn new material at the end of an oracle recorded at a previous session and reloaded for the
circumstance. So the long-term memory may become a very long-term indeed, including ses-
sions over months or years. One could even imagine that Omax learns the whole available
music played by a performer and thread its own improvised path through this database.

One must remember that, even if OMaxLisp can maintain several Oracles at the same time
(and improvise on them), it learns only in one unique Oracle, called MAIN.

In order to perform such oracle housekeeping, you will use the panel pictured above.
• Save to File saves the main oracle to a file (important convention: MyNameOracleFile.or)
• Load (red buttons) load an oracle saved from a previous session into a memory slot
• Wake up (blue buttons) the oracle loaded in the corresponding memory slot (including the
Main Oracle on the left) becomes the active improvising oracle
• Load Main (upper left white button) is similar to the red load buttons, that is it loads an
oracle from a file. But in this case, the loaded oracle becomes the MAIN, the oracle that
learns. Learning will occur at the end of the sequence contained in the loaded Oracle. This
is how you can augment a past session with the contents of a new session.

Notes:
• when an oracle is awakened to play, the interface in the improvisation control panel is up-
dated to reflect its characteristics (duration, number of phrases). The region mode is set to
All, the current region to (0,0) and the quality to 3
• When you load a MAIN oracle, the previous one is deleted so you loose your data except
if you saved it to a file beforehand
• You can have simultaneously at hand: a MAIN oracle and 8 loaded oracles among which
you can switch using the blue wake-up buttons
• When you switch from MAIN to another Oracle, the new oracle becomes the improvising
one, but the MAIN keeps on silently learning whatever input is played by the performer
(if record enable is on).

14
14
Other functions:
• Reset Agent resets the MAIN Oracle. This is comparable to OMaxInitSession, except that
you won’t loose all your data stored in the loaded oracles memory slots. It clears only the
MAIN (learning) Oracle. OMaxInitSession resets everything for a brand new session.
• Legato slices if not set, will play durations exactly as learned (i.e. if a chord is slightly ar-
peggiated, it will be played as is). If set it will quantify the duration (i.e. the chord will be
played with simultaneous onsets and offsets). This affects only MIDI mode, not audio.
• Short Endings when the performer stops playing, the oracle does not record the silence
(which could be very long) before he plays again. Rather it records a max 2 seconds pause
after the last event. If short endings is set, when the improvisation crosses such an ending
event, it will not play a 2 seconds pause. Rather, it will give the event the duration of the
preceding event.
• Debug light shows the recombination process in the OMaxLisp listener.
• Debug Heavy shows all the OSC messages flowing from OMaxLisp to Max.

About the debug mode

When using Debug Light, you will see in the Lisp window the recombination process. This
can help to tune up the Quality parameter (see Tutorial 2). When it says “No candidates”, it
means that OMax does not find any recombination and will pursue the original sequence until
it founds a possible jumping point. When it finds a recombination, it says “Jumping from x to
y” where x and y are positions in the oracle. In addition it gives the length of the context and
the rhythmical quality.
Too many “No candidates” may either show that the OMax improviser is currently in a par-
ticularly unique (non redundant) passage, but it might be that the quality factor is too restric-
tive (the value is too small, e.g. 1 or 2). Try a bigger value (e.g. 3).
Keep it mind that the debug printing takes processing power (on PPC).

15
15
Tutorial 5. Going Audio
In audio mode, all you have learned in Tutorials 1-4 stays valid, except now you’re learning
directly from an audio signal, and OMax improvisations are played in audio using the very
sound recorded from the performer.

After opening the OmaxMax patch:

• select the default audio instrument in the Initialization menu, press
OK. A new audio panel will appear.
• In the improvisation command window, select “normal”, or any size
according to the room available on your screen. A double waveform
display appears. The top (blue) large waveform displays a current
zoomed portion of the buffer. The small (yellow) always shows the
whole recorded buffer. By selecting a portion in the yellow pane, it is zoomed into the blue
pane. The all button redisplays the whole buffer in the blue pane. In order to select an im-
provization region, you have to click-drag in the blue pane, the yellow one being only for vis-
ual zooming purpose.

Let us review all the available parts of the interface:

• The top right blue is the

initialisation process.
• The orange is the input
panel, changing according
to the setup in te blue.
• Between them is the input
matrix, above the master.
•The green is the MIDI
window
•The navy blue is the
housekeeping window
(Tut. 4)
• The square beige yellow
is the segmentation win-
dow (Tut. 3)
• The hawaiian blue is the
output monitoring window

• The variable size impro-

visation control window
shows the waveform.
• The window also allows
monitoring for continuous
impro,control of the im-
provisation type (ALL,
Region or Follow, Tut. 2),
of the quality.

16
16
Check the audio on.

Check in the Max DSP options menu that every thing is set
up for audio input output. By default you should have 2 in-
puts: adc1 for recording the performer’s sound, adc2 for
pitch detection.

Setting up the recording

Two sound inputs are needed: one for recording itself (rec) and the other for pitch – or spec-
tral – detection (dtect).
A classical setup for e.g. a saxophon player would be:
• A good (best possible) aerial microphone on rec for recording a quality sound (this is
the sound you will hear when OMax wil be improvising)
• A contact microphone on the instrument on dtect (pitch detection is better with a close
take and you don’t want the pitch detection canal to take the surrounding sound, in-
cluding the OMax improvisation, because this would result in a feed-back process).
For pitch detection, a cheap piezo pickup on the reed is perfect; for spectral audio, the
pickup can be fixed on the bell.
An alternative would be to use a single close mike (e.g. a mike clipped inside the bell) and
send it to both rec and dtect. In this compromise, the sound for recording will be medium
quality and the detection, hopefully, will not catch too much of the sound environment.

In any case you can chose wich adc~ goes to which input:
• except in the mode MIDI and pianobar, both inputs (rec and dtect)
are needed.
• for pianobar mode, the dtect is disconnected. Only recording is
needed since detection is made by the MIDI device
• for MIDI, there is no sound and all this is irrelevant.

For test purposes, you can just use the Mac’s internal microphone (which is sent to adc1 and
2) and use a headset to avoid feedback.

Using a File as Input

It is perfectly possible to use a recorded
file as an input. Therefore, check the big
toggle written read file: it opens a win-
dow poetically called BOGUS_input.

17
17
 The use of this window is pretty straightfor-
ward:
• Use the open button to open a soundfile.
• You can play either by using the left toggle or the 1
and 0 buttons.
• You can play selected parts of the file by entering
their dates in ms. The entering of the left date (begin-
ning) forces the file to play
• You can pause, resume... during the reading of the
file,

Finally…
Proceed as in Tutorial 1: OMaxInitSession, Record Enable, and you’re gone.

For more details, check the sections

Ref1: Audio Control Panel Reference
Ref3: Recording setup

18
18
Tutorial 6. Sound Output Control
Monitoring
The Monitoring window allows to observe the sound busses. The MIDI activity is also visible
through the leds (it can be disabled by unsetting the MIDI toggle).

In order to understand the different signal-level meters in the Monitoring panel, here is the
OMax terminology:
• Continu the continuous improvisation generated by the active Oracle
• Blocks the (up to 5) sequences played using the Block panel (see Tut. 9)
• Seq the recorded sequence is played straight using the SeqDir Panel (see further)
• Dir the direct input signal as captured on the input rec
• Rec the record-to-file channels
• Cue the signal is exclusively routed to a cue bus (for example headphones), and its
normal output is muted. By default the output 9 is for the cue.
The numboxes above the meters allow for delays to be added.

Routing
Checking the Routing toggle opens up a sound
Matrix window.

The 9 signal tracks (continuous, block1-block5,

seq, dir, cue) can be routed to up to 9 audio out-
put (depending on your hardware setup) and/or
to 2 record channels (Left / Right).
By default, the direct signal (dir) is NOT sent to
the sound system (as normal for an acoustic in-
strument).

Recording
In order to record your mix, open a new sound
file using the Open button in the Monitoring
panel. Press the Start / Stop toggle to start/stop
the recording. By default the direct input dir IS
recorded, and push the toggle start/stop to start
the recording. WARNING: the recording is
quite demanding for the machine, on PPC.

19
19
Interlude 7. Audio Buffers
As previously said, there is almost no difference in manipulation between MIDI mode and
Audio mode, except for a few more controls in the last one. However, going audio involves
internal differences that have to be understood in order to comprehend the system’s behavior.
In the Midi mode, there is a model of the sequence played so far by the performer, maintained
by OMaxLisp, called an Oracle. This learning oracle is called the Main. Other Oracles can be
present at the same time, loaded from files. One can switch from one to the other, making it
the active improvising Oracle. Only the Main oracle learns from the real time input. One can
also load a Main oracle from a file, in order to augment it by learning from the real time input
(the new information is learned at the end of the one loaded).

Audio is in RAM
In audio mode, all this stays true. However, each oracle (Main or loaded) has an audio coun-
terpart in the form of a Max audio buffer. An oracle and its corresponding buffer are both se-
quences representing the same musical process that, at some time, has occurred, been re-
corded and analyzed. Oracles are symbolic sequences augmented with analytic structures,
whereas buffers are just plain audio buffers, but they represent the same musical sequence: in
particular, they have the same total duration.
Buffers are the result of recording the input audio signal on input rec, while Oracles are the
result of recording and analyzing the symbolic data send by OMaxMax to OMaxLisp. This
symbolic data is the result of continuously analyzing the input signal and extracting high level
features (for the time being, events boundaries, pitches, intensities or, in spectral mode, events
boundaries and spectral descriptors).
In the audio mode, every operation involving an Oracle involves and audio buffer as well. For
instance, loading an oracle from a file will load a soundfile into a buffer as well. Saving an
oracle will also save the corresponding buffer to a soundfile. When a 150 MB buffer (the
maximum, thus 20mn of sound) is involved,this is more time consuming than in simple MIDI.
Memory limitations will not allow you to do everything you were doing in simple MIDI. It’s
up to you to handle this, providing there is no limitation in OMax (even the 150MB max
buffer size can be changed if you have enough RAM). However, it is recommended to start
playing with not too large buffers if you use the load features. It also depends on your ma-
chine and, even more, on the size of your RAM. In any case, it is recommanded that you do
not learn ot improvize when you are loading files: you could get serious overload problems.

Saving / Loading in Audio Mode

When you save an oracle in audio mode (see Save in Tutorial 4), e.g. in file MySession.or,
OMax will automatically save a soundfile called Mysession.aiff in the same directory. This
soundfile is an image of the audio buffer corresponding to the Oracle saved. WARNING: the
“.or” termination is madatory. Both files have the exact same name with different endings. On
Intel version, avoid special characters (specially accents) in filenames.
When you load an oracle in audio mode (see Load in Tutorial 4), e.g. from file MySession.or,
there has to be a soundfile named Mysession.aiff in the same directory. This sound file will be
automatically loaded in order to fill the audio buffer homologous to the oracle.

20
20
Tutorial 8. Going Spectral
In spectral mode, all you have learned in Tutorials 1-7 stays valid; you’re now learning di-
rectly from a full audio signal spectrum, not from pitch detection, like in audio mode (Tut. 5).
The incoming spectral frames are turned into vectors of spectral descriptors through the FTM/
Gabor Library (this is why you need to install FTM in order to take advantage of the spectral
mode). Therefore there is no relevant MIDI information. However, like in audio mode, OMax
improvisations use the same audio sound recorded from the performer.

After opening the OmaxMax patch:

• select SPECTRAL_lpc in the Initialization menu, press OK. A new orange spectral panel
will appear.
• If the improvisation window is on “number”, that is, shows a timeline, select “normal”, or
any size according to the room available on your screen. A double waveform display appears.

•This is the way the orange panel looks:

• This has common featu-
res with the audio panel
(Tut. 5), notably the Gbf2
to check the waveform.
• It is possible to change
parameters and even des-
criptors (going back and
forth between mfcc and
lpc) from this window.
However, this should be
done ONCE before the
session.
• The detect level corres-
pond to the signal sent to the descriptors. A smaller value will permit to have detection of silences and therefore
“phrases” and stops in the recording process. Beware that by cutting too soon, the sound itself risks being stran-
gely cut.

Check the audio on.

Check in the Max DSP options menu that every thing is set up for
audio input output. By default you should have 2 inputs: adc1 for re-
cording the performer’s sound, adc2 for spectral analysis. Again a sin-
gle mike input can be routed to rec and dtect if you can cope with
feedback.

21
21
Spectral audio, quality and continuity
In MIDI and Audio modes, the events roughly corresponds to “notes”. This is actually the
case for a monophonic instrument. Each detected note is an event. For polyphonic instruments
like keyboards, events correspond more to points of attack (or release) of at least one note.
The spectral approach is radically different: like a sound-film, the sound is analysed in
frames, up to more than 80 per seconds (fortunately for the machine, OMax aggregates simi-
lar contiguous frames into “super-frames” and therefore does rarely have that many).
The values for the continuity factor are thus very different in spectral setup. Continuity of 8 or
20 notes from an original recording can be many. 8 or 20 spectral frames generally are not
even a second of sound.
A good rule of thumb is to think that spectral continuities are 8 to 10 times more than “event”
continuity values. Of course, there no direct relationship: one long note with vibrato could be
one event and many frames, while a very fast flute trill could generate almost as many events
as frames.
What is said for continuity goes of course for quality: the quality factor is the context quality ;
4 notes in common are a lot. 4 frames too, actually, because it means not only a timbre, but a
timbral evolution on something like 100 milliseconds. However, one should prefer (very)
good qualities in spectral environement. Where 3 is an acceptable quality in event-mode, 2 is
better in Spectral mode.

So quality = 2 and continuity = 60 are good values to start with.

Setting up Recording
As in Tutorial 5.

Input from File

As in Tutorial 5.

Finally...
Proceed as in Tutorial 1: OMaxInitSession, Record Enable, and you’re gone. Take care,in
spectral mode the recording generally starts as soon as the Record Enable button is set.

For more details, check the sections

Ref4: Spectral Control Panel Reference

Ref3: Recording setup

22
22
Tutorial 9. Block Impros and Loops
By checking the toggle block, you open a new window with a 5 tracks mixer-like interface.

This allows you to “silently” compute new OMax improvisations and store them as se-
quences (called Blocks) to be played whenever you like. A Block will always play the same
sequence, thus it is “improvised” upon first computation, but not every time you play it. Con-
stant pattern can then be kept in an otherwise always evolving system. Due to the concurrent
agent architecture in OMaxLisp, computing a block does not interrupt other processes going
on, such as Main “continuous” improvisation. Each time you compute a new block impro, it is
stored as a sequence into the next available checked track (see little check boxes on the left of
the tracks). Uncheck the track if you would like to protect it from being erased by a subse-
quent block computation and to be sure to be able to replay it later.
When computing/loading a block, you shouldn’t try to compute/load another block before the
process is completed. A yellow light indicates that computing/loading process is taking place,
followed by a blue light when the process is completed. A green light indicates a block previ-
ously loaded. So the blue light aways denote the last computed block.

To compute/load a block, type a duration in msecs in the lower-right numbox and press the
Block Impro button. A yellow light appears, wait until it turns blue. The Impro is computed
from the currently active Oracle, with the current Region/Quality/Start on Border parameters.
The BLImpro on phrase button computes a block starting on the indicated Phrase number,
with the current active Oracle set in mode All.

By using the loop mode and playing a polyphony of blocks, it is possible to get an extremely
dense texture., It is possible to loop one block, or all of them, to start a block immediately af-
ter it is computed (immed), to allow only one block to play at a time (1-block). Of course each
block can have its individual level adjusted, can be muted or send to a cue chanel (to be liste-
nend silently in headphones for instance)

23
23
In the following picture, indicators are in italic, while active controls are in regular font face.

24
24
Tutorial 10. Presets for Improvisation Control

The Improvisation Control window has a librarian module for storing control presets.

Its primary use is during a performance, to remember particularly successful or interesting

moments. The complete setup of the control window is remembered, that is: Quality, Region
mode (All, Region or Follow), Region boundaries (and whether they are Phrase or Time lim-
ited), possibly follow time (in ms), Bordr Strt (0 or 1) and the continuity value. However,
they can be saved in a file to be recalled with the corresponding Oracle. (By default, the file
#1_lj.xml present in the OMaxMax folder will be loaded at start). This is the role of the but-
tons write and read.
Any setup can be instantly stored with the Button New. The new setup will take the upper po-
sition in the menu. The current setup can be modified to the current situation (button Modify)
or erased (Erase).
The first setup of the menu [rien] has no action on the user interface.

The menu shows the value of the parameters in this order:

Region mode / Start on a border / Phrase or time limit / Limits / quality
• Region mode: A (=ALL)/ R (=Region) / FL (=Follow)
• Start on a border: [ (=yes) : (=no)
• Phrase limit: Ph (=yes) (nothing = no)
• Limits for ALL and Regions: XX-YY in seconds or in phrase numbers
• Time for FOLLOW: XXms
• Quality: Q1, Q2, Q3 or Q4
• Continuity: 1 to 248

Examples:
• R:43-54_Q3_c25 (Mode Region, from 43349 to 54289 millisec in the sequence, qual-
ity 3, continuity 25)
• A[Ph2-4_Q2_c8 (mode ALL, Start region Phrases 2 to 4, start on phrase begining,
quality 2, continuity 8)
• FL16235ms_Q3_c10 (mode Follow, follow time = 16235 msecs, quality 3 and conti-
nuity 10)

Notes on the use of the Library:

The Library is very much Oracle-dependant. Any setup may be irrelevant for a given oracle:
for instance, the phrase or time borders could not exist in the given sequence. In general,
OMaxLisp switches to a default value when the borders asked for are impossible to reach.

25
25
Punchin / punchout
Another way of specifying and storing a region for later use is the Punch-in Punch-out win-
dow. To open this window, check the punchin toggle.

By clicking on the Punch in/out toggle, a region is created on the fly and stored in the library.
This way you can specify a region in real time according to what you hear. The quality, start
on region border and phrase limits parameters associated to the punched region can be set
beforehand using this interface.

26
26
Ref1: Audio Control Panel Reference
Control panel

• SeqDir allows to directly play through the sound buffer, that is, to play back the re-
corded sequence as the performer played it. One sets the starting point and the length
of the excerpt (in millisecs). It can be muted, sent to the cue bus, and the level numbox
can be adjusted in dB.
• The vertical slider on the left indicates the progressive filling of the Main buffer
• The input meter is the input level of the main (recorded) signal on the rec channel
(usually adc1)
• The detect meter indicates the level of signal sent to to pitch detection (or spectral
analysis) from the dtect channel (usually adc2). The white led just at the right blinks
every time an event is recognized. Warning! A very low level signal is generally suffi-
cient for the pitch detection. This is what the level value is for (see also Ref3)
• Quality, window, proba: these are set automatically when you choose an audio in-
strument in the initialisation panel.
• Read file for learning from a soundfile instead of a real time input. Opens up a panel
where you load, start and stop a soundfile. This is a nice way of testing the system
with a good input quality.
• Listen MidiDtct sends the output of the pitch detection to the Midi Module. This is for
debug purposes. A better way to listen to the Midi detection is to turn on the Detection
toggle on the Midi control panel (see below).
• Listen Direct Signal sends the direct real time input from rec channel (usually adc1)
to the Sound Matrix. The vertical slider controls the level. It is advisable to let it on,
and to control this directly on the Sound matrix.
• Yin, Stat: debug and level purposes (see also Ref3).

Pitch Detection
In Audio mode, like in other modes, it is preferable to set the gen-
eral parameters before the start of the session. Any change of pre-
set (marked by the OK button) empties the buffer. The presets cor-
respond to existing instruments, and can often work with success
with others.
However, the values of the detection parameters can be adjusted on
the sound control panel.

Generally speaking:

27
27
 • The default values are a good starting point
• Some of these values, notably Sound-Poetry and Percussion, work for extremely com-
pact and dirty sound, working more as spectral enveloppes than real pitch detectors
• The window size (40-80ms) must be larger for slower and lower instruments
• The quality (0.35-0.90) must be lower for less harmonic (“dirtier”) sounds
• The probability (0.40-0.80) can be lower if the sounds are hard to detect.
• The level should be low enough to prevent the ambiant noises (including OMax own
improvisation) to trigger the detection. The detection can be monitored on the black
and white led.

Buffers
The buffers and their content are in the subpatch Gbf2 at the top of the panel. By clicking on
it, one gets acces to the content of the buffers.

At the beginning of the session, the default values for the length of the main buffer (20 min-
utes), the sample size (24bits) and the format (aiff) can be changed.
By double clicking in the buffer object, you can verify that the buffers acctually contain sound
(if not, check that Audio is on).

28
28
Ref2: MIDI Control Panel Reference
MIDI panel in MIDI mode... and in Audio mode

The green MIDI panel changes its appearance when switching from MIDI mode to Audio
mode. In both modes, one can set the MIDI velocity using the vertical slider. The big toggle
above the slider sets the MIDI output on or off. One can notice in the above pictures that by
default, the toggle is On in MIDI mode and Off in Audio mode.

MIDI mode
MIDI In subpanel (left side)
• MIDI port selection menu
• MIDI in channel filters (toggles)
• Noteout toggle (= through, bottom left) : sends the MIDI input to the output.
MIDI Out subpanel (right side)
• MIDI port selection menu
• Through/-127/fixedc chan: affects the MIDI channels of the OMax improvisation
output. Through: same channels as the input. -127: not used in MIDI mode. Fixed:
user defined single channel.
• Save: save the MIDI input and Output happening in a session to a file. It has to be
checked at the beginning of the session. At the end, upon unchecking, it will ask for a
filename ( .mid). In order not to create confusion between the input and output chan-
nels, there is an offset parameter. For example, if the input is on channel 1, 3, 5 and
the offset is 8, the Midifile will contain the performer’s input in channels 1, 3, 5, and
the corresponding OMax improvisation on 9, 12, 14.

Audio mode

29
29
In audio mode, two MIDI flows can be generated. The output of the pitch detection (Detec-
tion, right panel) which corresponds to the performer’s input, and the output of the OMax im-
provisation (MIDI Out, left panel).
The controls have basically the same meaning than in MIDI Mode. The default MIDI Channel
option is set to -127 because in audio, the event have a MIDI channel >= 128. So 128 is really
Midi channel 1. You shouldn’t change this option except if you would like to set a fixed MIDI
channel (e.g. 13).
The Detection toggle allows the MIDI output of the pitch detector to be played: the acoustic
instrument is doubled by its MIDI counterpart. The MIDI output channel is set in the numbox.

Spectral mode
MIDI flow is irrelevant and should not be used. Therefore the MIDI window disappears.

Ref3: Recording Setup Reference

Principles
There are two sound inputs to OMax audio. By default they are patched to adc1 and adc2
1. Recording itself (rec)
2. pitch – or spectral – detection (dtect)

Sending both inputs to the right place

• Except in the mode MIDI and Pianobar, you need an actual audio
input (adc 1 or 2) to be plugged to rec AND one to detect
• in pianobar mode, dtect is disconnected. Since detection is made by
the device attached to the instrument, only recording is needed
• for MIDI, there is no sound and all this is irrelevant.

A normal setup for e.g. a saxophon player would be:

• A (very) good aerial microphone on rec for recording a quality sound (the sound one
hears when OMax is improvising)
• A contact microphone on the instrument on dtect (pitch detection is better with a close
take and you don’t want the pitch detection canal to take the surrounding sound, in-
cluding the OMax improvisation, because this would result in a feed-back process).
For pitch detection, a cheap piezzo pickup on the reed is perfect; for spectral audio,
the pickup can be fixed on the bell.
An alternative could be to use a single mike very close to the instrument (e.g. a mike clipped
inside the bell) and send it to both rec and dtect. In this case, the sound for recording could be
OK and the detection, hopefully, will not catch too much of the sound environment.
For test purposes, you can just use the Mac’s internal microphone (which is sent to adc1 and
2) and use a headset to avoid feedback.

General recording Advice

For recording, best is the rule
The best mike in the best position close enough to the instrument is the best solution.

30
30
But for pitch and spectral detection, low-fi is the rule
Curious as it may be, a short bandwith is enough (and actually better) for a good pitch detec-
tion AND for spectral descriptors extraction. This should not be curious: although the record-
ings are terrible, we are perfectly aware of the quality of Caruso’s voice or Louis Armstrong’s
trumpet on their recording. The information given by a better bandwith (although extremely
nice when listening) is as enormous as superfluous, making a hard time for the analysis proc-
ess. This is the reason why a very cheap contact microphone will generally give better results
than a good one.

Adjusting the detection/spectral level

Audio mode
The detection level for the dtect signal is very important. It can be set in the
right part of the Audio control panel. WARNING: the level is set in the number
box. The slider on the right correspond to the level of the direct signal even-
tually sent to the output (if the output matrix is
on for direct sound)
By crossing the stat toggle, a window opens
where you can control the levels.
This is the windows where the statistics for pitch
detection are made.
The most important number is the one in the cen-
ter, written 207 on the figure. This number rou-
ghly corresponds to a MIDI velocity, that is:
• piano sounds should be between 10 and 50 (40
is a good value in case you risk feedback)
• fortissimo sounds should be between 120 and
140 (but this is less important).
• 207 is a special value corresponding to NO-
TEOFF
So the best way to set the level is to play a fortis-
simo sound and set the level to have something
between 120 and 140, then to adjust the level to
get something like 40 in pianissimo.
Finally, you should check that when OMax plays
fortissimo, it does not trigger any pitch detection.
If it is not the case, lower the level. If trouble
persists, use a contact microphone for the detec-
tion.

Spectral mode
In spectral mode, the recording is almost
ALWAYS on, as soon as you push the Re-
cord Enable button. However, you can get phrases and even stopping with a very silent con-
tact mike, but you should not count on it too much. To stop recording, putting Record Enable
to zero is the best solution. However, you can also use the level number in order to get phrases
and to stop the recording when nothing happens. It is advisable not to go too low: there is the
risk of getting unpleasant cuts into the sound events.

31
31
Ref4: Spectral Control Panel Reference
Spectral Oracle Basics
The spectral oracle uses what is called “spectral descriptors”. These are simple descriptions of
the sound spectrum. They are calculated by “frame”. We have two ways of transforming these
in oracle events: quantizing them, and, if the quantification of several successives frames gi-
ves the same result, we take them as one single longer frame. In any case, since the Oracle
works with an alphabet, we must be able to transfrom these descriptors values in disctinct
equivalence classes. Quantification is one way to get these classes, and also allows to get less
values for the alphabet: a spectral descriptor consisting of 20 coefficients varying between 0
and 9 would create an alphabet of 1020 possibilities. Chances to get repeated patterns be mi-
nimal! Another way of getting a shorter alphabet is, of course, to start with smaller descrip-
tors, that is, with less coefficients. The whole process of getting equivalence classes with
spectral description is still a very active field of research. Our use of “crude” quantification
and of a small number of coefficient is a very basic method, although relatively efficient.
The two descriptors now implemented are mel frequency cepstrum coefficients (mfcc) and
linear prediction coding (lpc). The quality of the results depends on the music played, of
course, but also on the choice of the coefficient and of the way they are quantized. One must
be aware that the idea behind the oracle is to find repetitions (only possible with a limited
number of possibilities) such that these repetitions must sound close enough that one can be
taken for the other.
More details on the spectral descriptors is available in the bibliography.

Control panel

• SeqDir allows to directly play through the sound buffer, i. e. to play back the recorded
sequence as the performer played it. One sets the starting point and the length of the
excerpt (in millisecs). Mute, Cue bus send and Level in dB can also be adjusted
• The read file button opens a window allowing to read from a file (see Tut. 5)
• The level number decides of the level for the sound going to the spectral analysis. It
should not be confused with the right slider, which decides what level of the direct
sound is be sent to the sound system for possible amplification.
• by default, the direct button is checked and the incoming sound is sent to the output
matrix (Tut 6). However, in this matrix NO output is selected. To actually amplify the
incoming sound, both the direct button must be on and an output chosen on the output
matrix.
• The Gbf2 subpatch allows to check the contents of the buffers and change the format
of the main buffer (see Ref1)

32
32
 • The change parameters allow to change the values of the spectral descriptors. Once
again, it is strongly advised to do it before recording a session. Before, however, you
can play with the values of the descriptors, even change of descriptor type with the
radio buttons.

Changing Spectral Parameters

For the moment, the best result we have are with the default parameters on lpc. However, for
very continuous sound (like some electronic sounds) mfcc are more efficient.

lpc
Parameters value:
• windowsize: as the name tells, the size of the window
on which the analysis is made. At 44.1kHz, a window of
4096 corresponds to 93 milliseconds and covers all pos-
sible frequencies.
• windowstep. More than a spectrum is calculated by
windowsize. Because of the use of windowing tech-
niques, most of the useful information is in the middle of
the window. Here, a spectrum is calculated every 512
samples (12 msec). A small value is more demanding to
the computer but permits a better time precision.
• lpc_bands. Given the characteristics of lpc, a uneven
value is preferable. Small means a cruder approximation
of the spectrum. If you get very strange recombinations
and have many of them, try 9 or 11
• quantization. It is a MULTIPLYING factor (in mfcc it is a dividing factor). The result
will be more precise with a higher value… and the spectra will be more differentiated
(you will get less recombinations with, normally, a better resemblance)
• set!!! or cancel nothing is done until one of these two buttons is banged.
The default values are a good start. It is very easy to get values that will not give any result :
15 bands and a quantization factor of 5 will give as many letters as frames and will result in a
total absence of recombination.

33
33
mfcc
Parameters value:
• windowsize: as for lpc.
• windowstep. as for lpc.
• lpc_bands. 7 is a relatively low number for mfcc. A
higher number could be tried (8 to 11)
•quantization. It is a DIVIDING factor (in lpc it is the op-
posite). The result will be more precise with a lower value;
the spectra will be more differentiated (you will get less re-
combinations with, normally, a better resemblance)
•set!!! or cancel nothing is done until one of these two but-
tons is banged.
The default values are a good start. A interesting procedure
is to compare two oracles made with the same recorded file
with slightly different values.

34
34
Ref5: SuperVP Control Panel Reference

The SuperVP panel is launched by dbl-clicking on the PitchTransf patcher, bottom-right part
of the OMaxMax patch.
When SuperVP processing is On for the continuous OMax improvisation or for one of the 5
blocks, the time structure and the pitches of this source stram are changed, depending on the
controls you have activated in the panel (see controls below).

The output of the transformed sound stream is sent to the Cue channel (see Tutorial 6 subsec-
tion Routing). Using the routing matrix, you can play the source improvisation, the trans-
formed improvisation, or both together. In the latter case, you might want to correct the delays
between the source stream and the transformed stream by adding a 10-50 ms delay to the
source stream (see Tutorial 6, Monitoring).

35
35
The same kind of transformations can be achieved on the direct instrument sound rather than
the OMax improvisation. In order to get a panel with similar operations on direct input, open
the helper patch DirTransf.

36
36
Ref6: Miscellaneous
Global Controls

Threshold of silence dura-

tion in ms for detecting Master Volume
phrase boundaries (Midi and Audio)
Note: the Pianoroll button is
disabled

Utilities

Clicking on the debug button in the seg-

mentation panel opens up this subpatch
where you can set the midi input note-off
segmentation threshold (default 15000
msec). It means that if no note-off has
occured in the midi input to stop note-on
that arrived more than 15secs before, and
no other, shorter note is pending, then the missing note-off are generated. This is to give an
end to ‘unfinite chords’ in the midi input, a phenomenon which occurs with unperfect midi-
devices (such as the Moog Pianobar or Midi Guitars) when note-off are eaten in the transmis-
sion and the performer stops playing for a while. You can change this value.

The patcher utilities opens up a window whe-

re you find the (very useful in Midi Mode) all
notes off panic button, a utility that informs
you about your IP address and a patcher that
opens the OMaxRemote automation messages
helper)

37
37
Automation Messages

This help patch documents the OMax automation messages. The red background contains si-
gnal messages (in and out) the blue one control messages (in and out). Using these messages
you can take full control of OMax using your own interfaces or external devices. Do not close
the OMax windows, just hide them if you don’t need them. For a single OMax utilization just
copy these messages to your patch. For a multi OMax utilization, the parameter #1 must be
changed to reflect the parameter sent to each instance of OMax (see Helper patches below,
MultiOMax), e.g. ‘s toto_improvise’ or ‘s titi_PTRegion 0’.

Helper Patches
You will find helper patches in the OMax folder.

PitchTransformations
With this patch, you will be able to perform on the fly transformations on the material gener-
ated by OMax:
• set-theoretical pitch transformation (MIDI mode only)
o Affine transform (a=2, 4, 8, 10 whole tone scales, a=3, 6, 9 diminished, a= 5, 7
permutation of the chromatic scale)
o Inversion: mirror inversion on the chromatic scale. The inversion with parame-
ter a = n is equal to the affine transform 11a+n
• Kbd inversion: maps notes that correspond to the center of the keyboard to the ex-
tremes and vice-versa (MIDI mode only)
• Tempo: slow down or accelerate tempo (2 means twice as slow, 0.1 means 10 times
faster)
• Retrograde: improvises the oracle backwards (works also in audio mode)

38
38
These transformations are set for the Midi mode. The retrograde works in audio mode, and
the tempo more or less.

MultiOmaxMax
You can launch up to three OMax sys-
tems hat will operate in parallel with
independent inputs and outputs but
without any interaction one with an-
other. In order to proceed, open the
helper patch MultiOMaxMax then
double-click on any subset of the three
patchers. The corresponding OMax-
Max patches will open up. On the lisp
side, you have to load the correspond-
ing OMaxLisp images. These are
numbered (OMaxLisp, OMaxLisp2
and OMaxLisp3) and correspond to the three OMaxMax patches in the order from top to bot-
tom. The OSC port numbers in the patchers arguments are mandatory since these are the ones
recognized by the corresponding Lisp Images.

39
39
Bibliography: OMax related Publications

Bloch, G., Dubnov, S., Assayag, G., «Introducing Video Features and Spectral Descriptors in
the Omax Improvisation System », Proceedings of the ICMC’08, Int. Comp. Mu-
sic Assoc., Belfast 2008.
Assayag, G. , Bloch, G. « Navigating the Oracle: a Heuristic Approach », Proceedings of the
ICMC’07, Int. Comp. Music Assoc., Copenhagen 2007.
Cont, S. Dubnov, G. Assayag « Anticipatory Model of Musical Style Imitation using Collabo-
rative and Competitive Reinforcement Learning », Anticipatory Behavior in
Adaptive Learning Systems, (Martin Butz and Olivier Sigaud and Gianluca
Baldassarre, Berlin), 2007
E. Amiot, T. Noll, M. Andreatta, C. Agon, Fourier Oracles for Computer-Aided Improvisa-
tion », ICMC 2006, New Orleans, 2006
G. Assayag, G. Bloch, M. Chemillier « OMax-Ofon », Sound and Music Computing (SMC)
2006, Marseille, 2006
G. Assayag, G. Bloch, M. Chemillier « Improvisation et réinjection stylistiques », Le feed-
back dans la création musicale contemporaine - Rencontres musicales pluri-
disciplinaires, Lyon, 2006
G. Assayag, G. Bloch, M. Chemillier, A. Cont, S. Dubnov « OMax Brothers: a Dynamic To-
pology of Agents for Improvisation Learning », Workshop on Audio and Music
Computing for Multimedia, ACM Multimedia 2006, Santa Barbara, 2006
Cont, S. Dubnov, G. Assayag « A framework for Anticipatory Machine Improvisation and
Style Imitation », Anticipatory Behavior in Adaptive Learning Systems (ABi-
ALS), Rome, 2006
Rueda, G. Assayag, S. Dubnov « A Concurrent Constraints Factor Oracle Model for Music
Improvisation », XXXII Conferencia Latinoamericana de Informática CLEI 2006,
Santiago, 2006
G. Assayag, S. Dubnov « Improvisation Planning and Jam Session Design using concepts of
Sequence Variation and Flow Experience », Sound and Music Computing 2005,
Salerno, 2005
Mondher, A. Gérard, M. Stephen, L. Olivier, C. Jean-Marc, R. Francis « De la théorie musi-
cale à l’art de l’improvisation: Analyse des performances et modélisation musi-
cale », ed. Mondher AYARI (DELATOUR-France, Paris), 2005
G. Assayag, S. Dubnov « Using Factor Oracles for machine Improvisation », Soft Computing,
vol. 8, n° 9, Septembre, 2004
S. Dubnov, G. Assayag, O. Lartillot, G. Bejerano « Using Machine-Learning Methods for
Musical Style Modeling », IEEE Computer, vol. 10, n° 38, Octobre, 2003
S. Dubnov, G. Assayag « Universal Prediction Applied to Stylistic Music Generation »,
Mathematics and Music. A Diderot Mathematical Forum, ed. Assayag, G.,
Feichtinger, H.G., Rodrigues, J.F. (Springer, Berlin), 2002
G. Assayag, G. Bejerano, S. Dubnov, O. Lartillot « Automatic modeling of musical style »,
8èmes Journées d'Informatique Musicale, Bourges, 2001
O. Lartillot, S. Dubnov, G. Assayag, G. Bejerano « Automatic Modeling of Musical Style »,
International Computer Music Conference, La Havane, 2001

40
40
G. Assayag, S. Dubnov, O. Delerue « Guessing the Composer's Mind: Applying Universal
Prediction to Musical Style », ICMC: International Computer Music Conference,
Beijing, 1999
S. Dubnov G. Assayag « Universal Classification Applied to Musical Sequences », ICMC:
International Computer Music Conference, Ann Arbor Michigan, 1998
J. Godet « Grammaires de substitution harmonique dans un improvisateur automatique »,
Paris 6 [DEA ATIAM], 2004
Laurier « Attributs multiples dans un improvisateur automatique », UPMC Paris 6 [DEA
ATIAM], 2004
Seleborg « Interaction temps-réel/temps différé », DEA ATIAM [Mémoire de stage], 2004
N. Durand « Apprentissage du style musical et interaction sur deux échelles temporelles »,
Paris 6 [DEA ATIAM], 2003
Poirson « Simulations d'improvisations à l'aide d'un automate de facteurs et validation ex-
perimentale », UPMC [DEA ATIAM], 2002
O. Lartillot « Modélisation du style musical par apprentissage statique: Une application de la
théorie de l'information à la musique », Paris 6/Ircam [DEA Atiam], 2000

41
41