Evanstein MAT Thesis SCAMP

University of California
Santa Barbara
SCAMP:
Suite for Computer-Assisted Music in Python
A project document submitted in partial satisfaction

of the requirements for the degree
Master of Science
in
Media Arts and Technology
by
Marc Paul Evanstein
Committee in charge:
Professor Curtis Roads, Chair

Professor Clarence Barlow
Lecturer Karl Yerkes
July 2019
SCAMP:
Copyright
c 2019
by
Marc Paul Evanstein
ii
To the Making and Breaking of Rules
iii
Acknowledgements
The work presented here owes a great deal to my colleagues and friends at the UC
Santa Barbara Departments of Music and of Media Arts and Technology. In particular,
I owe a special thanks to the members of my committee: Clarence Barlow, Curtis Roads,
and Karl Yerkes. The many unique perspectives on music and music-making that I
encountered at UC Santa Barbara were crucial to identifying and understanding the
musical problems addressed in this work.
On a personal level, the support of my family and friends, and above all of my
wife Emily, lies behind every significant undertaking I accomplish. Our cat Minuet also
deserves mention as the inspiration for the cute, yet roguish acronym ”scamp”.
iv
Abstract
SCAMP:
by
Marc Paul Evanstein
This document consists of two papers describing the SCAMP (Suite for Computer-
Assisted Music in Python) framework for music composition. The first — and more
substantial — of these papers outlines the framework as a whole, discussing its motivating
principles and design goals, stepping through the key features of its API, and situating
it within the context of other tools for computer-assisted composition. The second paper
goes into further detail about the sub-package of SCAMP for managing the flow of musical
time, entitled clockblocks.
The Code
Both SCAMP and clockblocks are hosted on PyPI (Python Package Index), where
instructions for installation and links to the source code can be found:
https://pypi.org/project/scamp/
https://pypi.org/project/clockblocks/
v
Contents
Abstract v
1 SCAMP:
A Suite for Computer-Assisted Music in Python 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Designed for Flexibility . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Introductory Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.1 ”Hello World” . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.3 Generating Notation . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.4 Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Details and Further Possibilities . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.1 Multi-Part Music . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.2 Multi-Tempo Music . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.3 Skipping Forward in Time . . . . . . . . . . . . . . . . . . . . . . 19
1.3.4 Envelopes and Continuous Parameters . . . . . . . . . . . . . . . 19
1.3.5 Playback Implementations . . . . . . . . . . . . . . . . . . . . . . 24
1.3.6 Additional Note Properties . . . . . . . . . . . . . . . . . . . . . . 26
1.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.4.1 Directions for Further Development . . . . . . . . . . . . . . . . . 28
1.4.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2 Clockblocks: A Pure-Python Library for Controlling Musical Time 31

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.2 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 A Simple Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.1 The Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
vi
2.3 Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.1 Parallel Clocks Example . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.2 Parallel Clocks Implementation . . . . . . . . . . . . . . . . . . . 38
2.4 Compensating for Calculation Time . . . . . . . . . . . . . . . . . . . . . 40
2.5 Nested Clocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5.1 Tempo Inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.5.2 A Nested Tempo Example . . . . . . . . . . . . . . . . . . . . . . 43
2.6 Comparison with other approaches . . . . . . . . . . . . . . . . . . . . . 46
2.7 Directions for Further Development . . . . . . . . . . . . . . . . . . . . . 47
2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Bibliography 50
vii
Paper 1
SCAMP:
A Suite for Computer-Assisted
Music in Python
This paper introduces SCAMP, a computer-assisted composition framework in Python

designed to bridge the gap between the continuous timing of synthesis-based frameworks
and the discrete timing of notation-based frameworks. SCAMP allows the composer to
quickly audition and iterate over musical ideas based on the sonic result, and then flexibly
quantize and export the music in western notation. SCAMP provides varied and highly
extensible utilities for playback, features easy playback and notation of microtonality
and glissandi, has a flexible clock system capable of coordinating multiple streams of
music following separate tempo curves, and can export notation in the form of either
MusicXML or Lilypond (via the abjad library). The goal of the framework is to address
pervasive technical challenges while imposing as little as possible on the aesthetic choices
of the user. For this reason, care has been taken to separate key elements of SCAMP’s
functionality into self-contained subpackages. This design, along with SCAMP’s many
output channels, allows a user to pick and choose the functionality they need and to
abandon the framework when it no longer serves the aims of a given composition.
1
SCAMP:
A Suite for Computer-Assisted Music in Python Paper 1
1.1 Introduction
1.1.1 Motivation
Consider a composer who wishes compose a piece for string quartet based on climate
data. Having downloaded the data in CSV format, they wish to process it and experiment
with different mappings by ear. Finally, having crafted their preferred mappings into an
overarching musical form, they wish to output some preliminary notation, and then
reshape the result by hand in their preferred score-writing software.
Or consider a composer who wants to write a piece for instrument and electronics
where a simple mass-spring simulation generates a stream of glissandi, which simultane-
ously drives a modular synthesizer and results in written notation for the instrumentalist.
Or consider a composer who wishes to write an algorithmically generated piano con-
certo in which the piano and orchestra follow separate accelerating and decelerating
tempo curves. The pianist requires a score notated from the point of view of the piano’s
tempo curve, while the conductor requires a score notated from the point of view of the
orchestra’s tempo curve, perhaps with a renotated piano part for reference.
What these scenarios have in common is the translation of musical data between
different domains. In particular, all three contend with the transition between the con-
tinuous domain of sounding music and the discrete (and idiosyncratic) domain of notated
music.
These considerations were the driving forces behind the creation of SCAMP (Suite
for Computer-Assisted Music in Python), a GPL3.0-licensed pure-Python framework for
music composition, which is the subject of this paper.
2
SCAMP:
1.1.2 Overview
Computer-assisted composition offerings can be broadly divided into two groups:

those aimed at the direct creation of sound, which usually treat time, pitch, and other
musical parameters as continuous (e.g. PureData [1], Max/MSP or SuperCollider [2]), and
those that aimed at the creation of a traditional western score (e.g. abjad [3], OpenMusic
[4], music21 [5]), which — due to the constraints imposed by notation — generally work
with time in a discrete way.
SCAMP is designed to bridge these two worlds. Instead of a score, the composer
interacts with an ensemble of virtual instruments, auditioning musical ideas in continuous
time, outside of notational constraints. Once this result is deemed satisfactory, the user
can then flexibly quantize the result and export it as music notation, in the form of either
MusicXML [6] or LilyPond [7] (via the abjad library).
Time in SCAMP is managed with a sophisticated system of clocks, capable of remain-
ing tightly coordinated while following arbitrary, intertwining tempo curves. Moreover,
nested relationships are possible; processes can spawn accelerating or decelerating sub-
processes. Although the concept of tempo suggests beats and fractional units of time, all
durations are in fact floating-point, defaulting to seconds if no tempo is indicated. When
notation is needed, it can be generated in reference to any of the clocks used.
SCAMP is fundamentally note-event-based; however, notes are very flexible objects,
capable of incorporating arbitrary continuous parameter curves. Microtonality, glissandi,
and continuous volume envelopes are simple to achieve. In fact, so long as the playback
method is designed to respond to them, arbitrary parameters can also be included and
shaped in a continuous fashion,
It should be noted that SCAMP is not a synthesis engine, instead offloading playback
to one of the following:
3
SCAMP:
• SoundFont-based [8] playback via pyFluidSynth [9]. (SCAMP handles the pitch-
bends and channel management required for microtonality and glissandi, hiding
such complications from the end user.)
• Playback using an external synthesizer via an outgoing MIDI stream. (SCAMP

likewise handles pitchbends and microtonality.)
• Playback using an external synthesizer via OSC [10]
• User-defined playback implementations created by implementing atomic playback

functions
The goal of the SCAMP framework is to act as a hub that connects the composer
to other resources, while offering as open-ended a mental model as possible. As such, it
does not aim to offer utilities, such as generative toolkits or scalar / harmonic models,
which align with particular aesthetic approaches. Such utilities may be imported from
third-party libraries, or may be constructed by the composer. (The scamp extensions
package also exists as a repository for such functionality.)
The basic features and limitations of SCAMP are summarized in Table 1.1.
1.1.3 Designed for Flexibility
Fig. 1.1 illustrates both the internal and external dependency structure of SCAMP.
The packages in yellow represent external dependencies. These dependencies connect
SCAMP to various forms of musical input and output: LilyPond notation (via abjad),
SoundFont interpretation and playback (via sf2utils [11] and pyFluidSynth [9]), MIDI
input/output (via python-rtmidi [12]), and OSC input/output (via pythonosc [13]).
The packages in blue together comprise the SCAMP suite. The scamp package itself
contains the main functionality of the suite, with time management (clockblocks), param-
4
SCAMP:
Features Limitations
• Compose in continuous time, quan- • Not designed for direct sound syn-
tize to notation in discrete time thesis
• Coordinate multiple time streams • Note based (though broadly)

using nestable clocks with
smoothly varying tempi • Generative toolkits, models of
scales, pitch class sets, etc. not in-
• Continuous pitch-space, glissandi, cluded in the main library
volume envelopes and other contin-
uous parameter curves • Timing is not sample-accurate
(does not run on the audio thread)
• Playback using SoundFonts, MIDI
stream, or OSC. User-defined im- • Not designed for live-coding
plementations also possible. (though improving live input is a
future goal)
• Output notation via Lilypond or
MusicXML • No graphical front-end
• Easily understood syntax, with sig-

nificant flexibility hidden behind
sensible defaults
Table 1.1: Key features and limitations of SCAMP
eter shaping and control (expenvelope), and MusicXML export (pymusicxml) separated
out into self-contained packages. Note that clockblocks relies on expenvelope, since the
tempo curve of a clock derives from the Envelope class.
This modular structure, inspired by the Unix Philosophy, was chosen for two key
reasons:
• Easier maintenance: keeping different aspects of functionality separate keeps the

scope and upkeep of each piece more manageable.
• Discardability and reusability: not all composers will need all of the functionality
that SCAMP has to offer. For instance, some may be interested only in polyphonic
5
SCAMP:
pyfluidsynth rtmidi
midi input and
SoundFont playback
output streams
sf2utils
interpreting
abjad SoundFont files
LilyPond output pythonosc

sending OSC
P
M
CA scamp
S
Main Functionality
clockblocks pymusicxml
Time Management MusicXML Export
expenvelope
Parameter Management
Figure 1.1: Internal and external dependencies of SCAMP
tempo control or direct MusicXML export. By isolating these components from the
rest of SCAMP, composers with different goals can potentially incorporate them
into other frameworks.
This second point is an important one: every composer develops a unique workflow,
reflecting their unique set of aesthetic concerns. For most, this ultimately means patching
together a variety of tools. When tools are bundled together, this patching process can
become cumbersome. SCAMP’s modular structure allows composers to use (and perhaps
repurpose) only that which is relevant to them.
One final aspect of SCAMP that provides enormous flexibility for the composer is
that it is situated within the broader Python ecosystem. This affords:
• Easy access to a wide variety of libraries that can be adapted for musical purposes,
such as data processing (e.g. numpy, sci-py, matplotlib) and machine-learning (e.g.
tensorflow, scikit-learn) toolkits.
6
SCAMP:
• In particular, access to the many forms of data input and output available through
both standard Python and third-party libraries.
Taken together, the goal of all of these design choices to create a framework that is
as adaptable as possible to the needs of different composers and compositions.
1.2 Introductory Examples
1.2.1 ”Hello World”
As an introduction to the SCAMP API, we begin with a program that plays a short
arpeggio:
# import the scamp namespace

from scamp import *
# construct a session object

s = Session()
# add a new violin part to the session

violin = s.new_part("Violin")
# looping through the MIDI pitches

# of a C major arpeggio...
for pitch in [60, 64, 67, 72]:
# play each pitch sequentially
# with volume of 1 (full volume)
# and duration of half a beat
violin.play_note(pitch, 1, 0.5)
The result of this program will be the sound of a violin playing a C major arpeggio, at
max volume, over the course of two seconds. The first step, after importing the SCAMP
namespace, is to create a Session object. Most SCAMP programs will start in this
7
SCAMP:
way, as the Session object is the central hub through which most of the functionality of
SCAMP flows. In the same way that a session in a DAW encompasses tracks, transport,
and recording functionality, the Session object in SCAMP inherits from — and thereby
combines the functionality of — an Ensemble, a master Clock, and a Transcriber (see
Figure 1.2).
Ensemble
Hosts and manages shared
settings and resources
for ScampInstruments
Clock
Session
Manages and coordinates
multiple streams of time
at different tempi
Transcriber
Records notes played
to a Performance object
for later reuse or notation
Figure 1.2: The combined functionality of a Session object
The Ensemble functionality of the Session object — i.e. its role as a host of the
instruments being used — is employed in the third line, where a new violin part is
created and stored in the violin variable. By default, new parts created in this way will
play back via pyfluidsynth using a General MIDI Soundfont, with fuzzy string matching
being used to find an appropriate preset based on the name of the part. Finally, the call
to play note takes three arguments: MIDI pitch (where 60 = Middle C), volume (on a
scale from 0 to 1), and length in beats.
1.2.2 Duration
One might reasonably ask at this point how long a beat is. As mentioned above,
Session inherits from the Clock class, which is defined in clockblocks, a subpackage of
8
SCAMP:
SCAMP detailed in a separate paper [14]. When a Session object (or Clock object
more generally) is constructed, it becomes the default clock used on the current thread,
and is also assigned a default tempo of 60 BPM, or one beat per second. By way of
these defaults, when play note is called in above example, it looks to see which clock is
operating on the current thread, finds the session (s) we created, and plays the note for
0.5 beats on that clock (which corresponds to 0.5 seconds at the default tempo of 60).
In this way, the whole four-note arpeggio lasts two seconds.
We can play back the arpeggio faster or slower simply by setting the session’s tempo
attribute. For instance, if we add the following line directly after the second line in the
above example, the arpeggio will last twice as long:
s.tempo = 30
Calls to play note are blocking by default, not moving on to the next line until the
note has finished. This conforms to musical expectations (playing a note takes time), and
is a model of musical time similar to that used by the Euterpe language [15]. However,
SCAMP also allows the user to start and end notes manually. In fact, internally, a call
to violin.play note(60, 0.7, 1.5) is essentially equivalent to:
# start playing a middle C with volume 0.7

# returns handle for further manipulation
note = violin.start_note(60, 0.7)
wait(1.5)
note.end()
(Note that, like the play note function, the wait function captures the current clock
from context and takes a number of beats as an argument.)
Finally, in addition to the options above, it is also possible to call play note in a non-
blocking manner by setting the ”blocking” keyword argument to False. For instance,
9
SCAMP:
the following would play two notes, each lasting two seconds, that overlap by one second:
violin.play_note(60, 1, 2, blocking=False)
wait(1)
violin.play_note(64, 1, 2)
It is important to note that durations do not have to be of rational lengths. For

instance, after constructing the session and adding a violin part, the following would
repeatedly loop through our arpeggio using random (floating point) durations between
0.5 and 1.5 seconds:
# import the random module from

# the python standard library
import random
while True: # loop forever

for pitch in [60, 64, 67, 72]:
# pick a duration between 0.5 to 1.5
dur = random.uniform(0.5, 1.5)
# play a note of that duration
violin.play_note(pitch, 1, dur)
The default tempo of 60 BPM was chosen because it allows the composer to think
in terms of durations in seconds if so desired. This exemplifies a guiding philosophy
of the SCAMP framework: musical parameters are treated as continuous during the
compositional process unless explicitly quantized. It is only later on, when converting
the music to a score, that quantization becomes a necessity.
1.2.3 Generating Notation
In order to generate notation from the above examples, we use the third function of
the Session object: its role as a Transcriber. The purpose of a Transcriber is to keep
10
SCAMP:
track of any notes played by the instruments that have registered with it, and to save
the result as a Performance, which is essentially a note-event list. This performance
can then be quantized and converted into a Score, which is capable of saving either to
LilyPond (via the abjad library) or to MusicXML (via pymusicxml, a part of SCAMP).
The advantage of first transcribing the music as a Performance and then converting it to
a Score is that the performance exists in continuous time and parameter space, outside
of notational constraints. In fact, Performances can be replayed, either in part or in
whole, with the same instruments or with different instruments, at the same tempo or
at a different or changing tempo, and can even be re-transcribed after such alterations.
To transcribe the first example above and convert it to music notation, we would do
the following:
from scamp import *
s = Session()
violin = s.new_part("Violin")
# begin transcribing (defaults to using

# all instruments within the session)
s.start_transcribing()
for pitch in [60, 64, 67, 72]:

violin.play_note(pitch, 1, 0.5)
# stop transcribing and save the

# note events as a performance
performance = s.stop_transcribing()
# quantize and convert the performance to

# a Score object and open it as a PDF
# (by default this is done via abjad)
performance.to_score().show()
11
SCAMP:
a) b)
c)
d)
Figure 1.3: Examples of notation generated from introductory examples. Snippets

1) and 2) result from steady note lengths. Snippets 3) and 4) result from random note
lengths, with 4) using a simpler QuantizationScheme.
This results in the notation shown in Fig. 1.3a. When show is called, the score
representation within SCAMP is converted to an abjad score, which then outputs and
compiles LilyPond code and displays the result as a PDF. It is also possible to instead
call show xml, which uses pymusicxml to export a MusicXML document and opens the
result in a score editor (e.g. MuseScore, Sibelius, Finale).
By default, the to score function quantizes the music to 4/4 time. If a different time
signature is desired, one merely has to provide it as an argument:
performance.to_score("3/8").show()
This results in Fig. 1.3b. A looping or non-looping list of time signatures can also be
provided, as can a list of beats on which to place bar lines.
1.2.4 Quantization
Things get more interesting when we use random floating-point durations, since some
quantization is required:
12
SCAMP:
# ---- setup code omitted for brevity ----

for _ in range(2): # loop twice

for pitch in [60, 64, 67, 72]:
dur = random.uniform(0.5, 1.5)
violin.play_note(pitch, 1, dur)
performance.to_score("3/4").show()
This results in Fig. 1.3c. By default, divisors of the beat up to 8 are allowed, which
may be more complex than desired. Simpler results can be achieved by using a custom
QuantizationScheme that limits the max divisor to 4:
performance.to_score(
QuantizationScheme.from_time_signature(
"3/4", max_divisor=4
)
).show()
This results in Fig. 1.3d. Considerable customizability of QuantizationScheme is

possible: not only is it possible to set the max divisor, but it is also possible to specify
the degree of preference for simpler ratios, or even to provide a list of all allowed divisors.
Moreover, while quantization is always done at the beat level, the beat length and divisor
preferences can vary from beat to beat, according to the scheme imposed of the user.
As illustrated in Fig. 1.4, the choice of quantization for each beat is made by compar-
ing the (floating-point) note onset and termination times that fall within that beat to the
grid formed by each divisor. The winning divisor is that which minimizes the (weighted)
square error. To the degree that a preference for simple divisors is specified by the user,
this square error is weighted by the indigestibility of the divisor (as described by Barlow
13
SCAMP:
in [16]). The user can also specify a relative weighting of onset vs. termination error.
Total Total
NOTE NOTE NOTE
DIV. Onset Term.
OFF ON OFF
Error Error
2
3
4
5
6
Figure 1.4: Illustration of the quantization process.
One weakness of this approach is that it does not allow nested tuplet structures.
A possible future development, therefore, would be the incorporation of the Q-Grid
approach proposed by Nauert [17], which allows for such structures, taking into account
their degree of complexity.
1.3 Details and Further Possibilities
We now consider in more depth the range of possibilities for playback and notation
that SCAMP offers.
1.3.1 Multi-Part Music
The above examples, for simplicity, were all single-part; however, writing multi-part
music is straightforward:
14
SCAMP:
from scamp import *

import random
# one way of setting an initial tempo

s = Session(tempo=100)
oboe = s.new_part("oboe")
bassoon = s.new_part("bassoon")
# define a function for the oboe part

def oboe_part():
# play random notes until we have
# passed beat 7 in the session
while s.beats() < 7:
pitch = int(random.uniform(67, 79))
volume = random.uniform(0.5, 1)
length = random.uniform(0.25, 1)
oboe.play_note(pitch, volume, length)
# end with a note of exactly the right
# length to take us to the end of beat 8
oboe.play_note(80, 1.0, 8-s.beats())
# define a function for the bassoon part

def bassoon_part():
# simply play quarter notes on random
# pitches for 8 beats
bassoon.play_note(
random.randint(52, 59), 1, 1
)
# start the oboe and bassoon parts as
# two parallel child processes
s.fork(oboe_part)
s.fork(bassoon_part)
15
SCAMP:
# have the session wait for the child

# processes to finish (return)
s.wait_for_children_to_finish()
# here we call show_xml, which will open
# the result in, for instance, MuseScore
performance.to_score().show_xml()
This results in the notation seen in Fig. 1.5. The fork method of the Session
object, inherited from the clockblocks Clock class, takes a function as a parameter and
runs it as a parallel child process. There is considerable flexibility here: child processes
can spawn their own child processes, and so on. Unlike a naive approach to parallelism
using Python’s built-in threading module, all processes forked in this way will remain
tightly coordinated, with the clock also compensating for calculation time.
Figure 1.5: Example of notated multi-part music in SCAMP, having been exported
as MusicXML and imported into MuseScore.
1.3.2 Multi-Tempo Music
One of the most powerful aspects of SCAMP’s approach to time-management is that

each forked processes is associated with its own Clock, and can have its own smoothly-
varying tempo:
16
SCAMP:
from scamp import *
s = Session()
trumpet = s.new_part("trumpet")
trombone = s.new_part("trombone")
# Have the session as a whole speed up to

# 100 BPM over the first nine beats
s.set_tempo_target(100, 9)
# here we define a trumpet part to run as

# a child process. The clock argument in
# the function definition gives us access
# to the clock this child process runs on
def trumpet_part(clock: Clock):
# play eighth notes for three beats
trumpet.play_note(67, 1, 0.5)
# tell the clock for this child process
# to slow down to 1/2 speed over six
# beats in the parent process
clock.set_rate_target(0.5, 6, duration_units="time")
# keep playing eighth notes until 12
# beats pass in the parent process
trumpet.play_note(67, 1, 0.5)
# Fork the trumpet part as a child process

# (It will be influenced both by its own
# tempo and that of the session as a whole)
s.fork(trumpet_part)
# Play quarter notes for 12 beats

trombone.play_note(60, 1, 1)
17
SCAMP:
# Stop transcribing and show the result

performance.to_score("3/4").show_xml()
This results in the notation shown in Fig. 1.6a. Within the context of a session-wide
acceleration from 60 BPM to 100 BPM, the clock for the trumpet part slows down to
half speed. (Note that, in SCAMP, “rate” and “tempo” are alternate names for the same
underlying property, just with different units: a tempo of 60 BPM corresponds to a rate
of 1, and a tempo of 120 BPM corresponds to a rate of 2, etc. Tempo units are valuable
for their musical connotations, while rate units are generally more comprehensible in the
context of nested clocks.)
a.) The Rubytles
 43                        
 = 60.0 ( = 65.1) ( = 70.6) ( = 76.6) ( = 83.1) ( = 90.0)  = 100.0
accel. ( = 97.3)
7
trumpet
 43
           
trombone
b.)
Figure 1.6: The same music quantized to two different clocks, having been exported
as MusicXML and imported into MuseScore.
An important aspect of notating polytempo music in SCAMP is that performances

can be transcribed in reference to whichever clock is desired. For instance, in the example
above, we can generate a score relative to the tempo of the trumpet part by replacing
the corresponding lines with:
18
SCAMP:
# fork returns the Clock associated

# with the newly forked process
trumpet_clock = s.fork(trumpet_part)
s.start_transcribing(clock=trumpet_clock)
This results in the notation shown in Fig. 1.6b.

Further detail on polytempo usage of clocks can be found in [14].
1.3.3 Skipping Forward in Time
It may be that a composer, having devised an algorithmic process lasting 20 minutes,

would like quickly test what this process sounds like at the 15-minute mark. In SCAMP,
this is possible by simply calling a “fast-forward” method on the master clock (generally
the Session object):
s.fast_forward_to_time(900)
Similar to the process of “nesting” described by Smoliar [15], this causes the program
to execute as rapidly as possible up to the appointed time, while skipping note playback
calls. In this way, the program is in exactly the same state as it would have been had it
run normally.
1.3.4 Envelopes and Continuous Parameters
The accelerandi and decelerandi in section 1.3.2 are an example of the continuous
shaping of a parameter over time, in this case tempo. As touched on in section 1.1.3, each
Clock manages it tempo using a TempoEnvelope, a subclass of Envelope, the piecewise-
exponential envelope class defined in the separate package expenvelope.
Instances of Envelope are used broadly in SCAMP when any musical parameter is
19
SCAMP:
changing over time, such as in glissandi or dynamic playback. They can also be employed
by the composer for the larger-scale shaping of algorithmic processes.
A simple, evenly-spaced Envelope can be constructed and plotted as follows:
e = Envelope.from_levels([60, 72, 66, 70])

e.show_plot("Evenly Spaced")
This results in the plot shown in Fig. 1.7 (plotting is done using matplotlib).
Evenly Spaced
72
70
68
66
64
62
60
0.0 0.2 0.4 0.6 0.8 1.0
Figure 1.7: A simple Envelope with even spacing.
Each segment of the envelope can furthermore be assigned a duration and a shaping
attribute:
e = Envelope.from_levels_and_durations(
[60, 72, 66, 70], [3, 1, 1],
curve_shapes=[2, -2, -2])
e.show_plot("Uneven with Shaping")
This results in the plot shown in Fig. 1.8. The values in the curve shapes attribute
range from −∞ to ∞, with 0 being linear, negative values corresponding to early change,
20
SCAMP:
positive values corresponding to late change.

Uneven with Shaping
72
70
68
66
64
62
60
0 1 2 3 4 5
Figure 1.8: An Envelope with varied segment duration and shape.
The Envelope class is equipped with a range of utilities, including integration (needed
by TempoEnvelope to determine the length in time corresponding to a certain number
of beats at a changing tempo), approximation of arbitrary functions, and mathematical
operations, such as addition, multiplication, division, and concatenation.
Glissandi
The Envelope in the example above can be interpreted as a glissando by simply

passing it to the pitch argument of the play note function:
instrument.play_note(e, 1.0, 4)
This results in the notation shown in Fig. 1.9. Note that the glissando is scaled to
the length of the note, and that, by default, the pitch of the glissando is renotated on
every beat and at every local maximum or minimum of the curve. In this case, since the
segment durations are in the proportions 3:1:1, this results in quintuplets.
21
SCAMP:
Figure 1.9: Glissando notation resulting from passing an Envelope as the pitch argu-
ment to play note, as compiled by LilyPond, via abjad.
Since directly creating an Envelope object each time a glissando is desired is would
be cumbersome, any list given as the pitch argument to play note is interpreted as an
envelope:
# results in evenly spaced glissando

instrument.play_note([60, 70, 55], 1.0, 4)
If segment durations and curve shapes are desired, a list consisting of [values, dura-
tions, curve shapes] may be used:
# results in segment durations of 2 and 1

# and curve shapes of -2 and 0
instrument.play_note([[60, 70, 55], [2, 1], [-2, 0]],
1.0, 4)
Microtonality
The reader may have noticed quarter-tone accidentals in Fig. 1.9. Microtonality in
SCAMP is as simple as using floating-point values for pitch. When using SoundFont-
based playback via pyFluidSynth (which utilizes the MIDI protocol) or a MIDI stream to
an external synthesizer, SCAMP internally handles all pitchbend messages. Since these
messages are channel-wide, notes that change pitch (or might change pitch) are placed
on separate channels, with channels being recycled automatically.
If more specificity of pitch is desired in the notation, the user simply has to turn on
22
SCAMP:
microtonal pitch annotations in the engraving settings:
engraving_settings.show_microtonal_annotations = True
piano.play_chord([62.7, 71.3], 1.0, 1)
piano.play_chord([65.2, 70.9], 1.0, 1)
piano.play_chord([71.5, 74.3], 1.0, 1)
This results in the notation shown in Fig. 1.10.
Figure 1.10: Microtonal pitch annotations.
Dynamics
An Envelope object (or its corresponding list shorthand) may also be used to apply a
continuous volume curve to a note. For instance, a forte-piano-crescendo can be achieved
in the following way:
fp_cresc = Envelope.from_levels_and_durations(
[1.0, 0.3, 1.0], [0.1, 0.9]
)
piano.play_note(60, fp_cresc, 4)
Since dynamic notation is more subjective than pitch notation, arbitrary parameter
curves like this are not currently notated. This is an area for future development.
23
SCAMP:
Anatomy of a play note call
As mentioned in Section 1.2.2, play note internally breaks down into separate calls
to start note, wait, and end. With continuous changes of parameter, further calls are
involved. For instance, consider the following:
# glissando and crescendo over 4 beats

instrument.play_note([60, 70, 55], [0.3, 1.0], 4)
This is equivalent to the following separate calls:
note = instrument.start_note(60, 0.3)

# start a 2 beat pitch-change ramp
note.change_pitch(70, 2)
# start a 4 beat volume-change ramp
note.change_volume(1.0, 4)
wait(2)
# start the second pitch ramp
note.change_pitch(55, 2)
wait(2)
note.end()
While a call to play note is clearly more succinct, there are situations where the
length and course of the note are not known in advance, and which are therefore better
suited to the second approach. This is especially true with the incorporation of live input.
1.3.5 Playback Implementations
In the preceding examples, all playback is done via pyfluidsynth using a default Gen-
eral MIDI soundfont. However, one of the major flexibilities built into SCAMP is its
variety of available playback implementations, as well as the ability to define custom
implementations.
24
SCAMP:
Consider the following setup:
s = Session()
# Calling "new_part" results in a default

# SoundfontPlaybackImplementation, and
# add_streaming_midi_playback gives this a
# MIDIStreamPlaybackImplementation as well
# (here, port 2 is used for output)
piano =
s.new_part("piano").add_streaming_midi_playback(2)
# "new_osc_part" initializes the instrument

# with an OSCPlaybackImplementation
synth = s.new_osc_part("synth", ip_address="127.0.0.1",
port=57120)
# Calling "new_silent_part" results in an

# instrument with no PlaybackImplementation
silent = s.new_silent_part("silent")
As illustrated by Fig. 1.11, session s will be set up with contain an ensemble of

three ScampInstruments named “piano”, “synth”, and “silent”. The “piano” instru-
ment is created with the default SoundfontPlaybackImplementation to which we add
a MIDIStreamPlaybackImplementation. By calling new osc part, the “synth” instru-
ment is created with an OSCPlaybackImplementation instead of the default. Finally,
the “silent” instrument is created with no PlaybackImplementation at all.
When a note is played on an instrument, each of its PlaybackImplementations is told
to start a note, do any pitch / volume / parameter changes, and then end the note at the
appropriate time. Thus, when a note is played using “piano” above, MIDI messages will
be sent to fluidsynth to generate sound, as well as sent out of port 2. When a note is played
using “synth”, appropriate OSC messages (with address patterns like ”synth/start note”
25
SCAMP:
SoundfontPlaybackImplementation
piano
(ScampInstrument) MIDIStreamPlaybackImplementation
s synth OSCPlaybackImplementation
(Session) (ScampInstrument)
silent (no PlaybackImplementation)

(ScampInstrument)
Figure 1.11: Illustration of playback implementations.
and ”synth/change pitch”) will be sent to port 57120. Finally, when a note is played
using “silent”, nothing will happen, since it has no PlaybackImplementations. (Silent
instruments can nevertheless be useful: for instance, as a reference, one might want
to notate the pitch collection being used by an algorithmic process without generating
sound. Or one could use a silent instrument to aggregate the notations of several sounding
instruments using different SoundFont presets, which might come in handy for, say, a
violin part using multiple bow techniques as well as pizzicato.)
1.3.6 Additional Note Properties
The play note function in SCAMP can take a fourth, optional properties argument.
This argument acts as a wildcard, accepting a properly formatted string or a dictionary
specifying a variety of playback and notation details:
# passing comma-separated key value pairs

piano.play_note(60, 1, 1,
"notehead: x, articulation: staccato")
# just a value; articulation type inferred
piano.play_note(60, 1, 1, "staccato")
# "play_chord" can take multiple noteheads
piano.play_chord([60, 65], 1, 1, "noteheads: x/circle-x")
# passing a dictionary also possible
26
SCAMP:
piano.play_note(60, 1, 1, {"articulations": ["tenuto",

"accent"]})
This produces the notation shown in Fig. 1.12. In addition to modifying the notation,
articulations like staccato and tenuto will affect playback duration, and accents will affect
volume. Exactly how these notations affect playback is an adjustable (and disableable)
setting with sensible defaults.
Figure 1.12: Illustration of notation resulting from various properties
The properties argument may also be used to determine note spelling, either directly
or by suggesting a key:
# spell the note with a sharp

piano.play_note(63, 1.0, 1.0, "#")
# play some notes in F minor

pitches = [65, 63, 61, 60]
durations = [1/3] * 3 + [1]
for pitch, dur in note_info:
piano.play_note(pitch, 1.0, dur, "key: F minor")
This results in the notation shown in Fig. 1.13.
Figure 1.13: Notes spelled both directly and by specifying a key context
The properties argument can also be used to define aspects of playback other than
27
SCAMP:
pitch and volume, if implemented by the PlaybackImplementation. For instance, with

a properly configured OSC receiver, the following will produce a 3 beat glissando with
increasing vibrato frequency and depth:
# on the other end, an OSC receiver is

# setup to play notes that take vibrato
# OSC messages as well as the usual
vib = s.new_osc_part("vibrato_inst", 57120)
# any property entries starting or ending

# with "param" will be treated as extra
# playback parameters
vib.play_note(
[72, 64], 0.5, 3,
{"vib_depth_param": [0.5, 2],
"vib_freq_param": [3, 8]}
)
Any keys in the properties argument starting with the prefix "param " or
ending with the suffix " param" are interpreted as additional playback parame-
ters. In the case of an OSCPlaybackImplementation this results in outgoing OSC
messages with address patterns “vibrato inst/change parameter/vib depth” and “vi-
brato inst/change parameter/vib freq”. As with dynamics playback, translating this to
some form of notational output is an area for future development.
1.4 Conclusions
1.4.1 Directions for Further Development
As mentioned above, some form of notation for dynamics and other user-defined
playback parameters is needed. For dynamics, a distinction will need to be made between
28
SCAMP:
note-level dynamics (such as sfz or fp) and phrase-level dynamic spans (such as cresc.
or dim.). The easiest approach will likely be to allow specific dynamic notations such
as these to be defined and applied to notes or phrases, influencing both notation and
playback, as opposed to trying to infer notation from playback (which is the general
approach taken in SCAMP). For other parameters, some form of graphical notation
incorporating the shape of the envelope curve may be appropriate. In general, since the
Performance class retains continuous-time musical data, alternate mappings to different
kinds of score are possible.
This points to the value of expanding the options for translating musical data. Cur-
rently, a Performance can be converted to a Score, but not vice-versa. Likewise, a Score
can be translated to an abjad/LilyPond or MusicXML representation, but not vice-versa.
One development goal is to enable two-way translation between all of these representa-
tions. Also, although SoundFont playback and streaming MIDI output are possible, it
is not currently possible to convert a Performance to a MIDI file, or to directly save
playback to a sound file. These are planned improvements.
In terms of playback, a medium-term goal is to offer SFZ-file-based playback using
LinuxSampler [18]. As a modern, open standard for sampled sound playback, SFZ files
offer exactly the kind of flexibility that SCAMP is designed to afford [19]. Another
playback goal is to enable certain note properties, such as “pizzicato”, to trigger a preset
switch, much as they do in professional score-writing software, such as Sibelius. Finally,
MAX/MSP and SuperCollider abstractions are currently being developed for receiving
OSC output from SCAMP.
One last area of development is in the domain of live input. Although SCAMP is
not designed for live coding, the ability to receive real-time MIDI and OSC input, and
to thereby shape the result of an algorithmic process, would be of great value to many
composers’ workflow.
29
SCAMP:
1.4.2 Evaluation
Although all of the planned developments above are worthy goals, the most important
next step is to create music with SCAMP and to cultivate a broad and varied user-base.
This will allow any future development to be guided by the true obstacles that composers
face while using this framework.
To this end, I plan to lead workshops and classes devoted to exploring computer-
assisted composition using SCAMP, as well as to develop detailed online documentation
and tutorials. SCAMP was developed and re-designed over a period of many years,
based on my own compositional practice. Its true potential, however, can only be discov-
ered through interaction with other composers, approaching it with potentially radically
different compositional aims.
30
Paper 2
Clockblocks: A Pure-Python Library

for Controlling Musical Time
This paper describes clockblocks, a GPL3.0-licensed pure-Python library for controlling

the flow of musical time, which is part of a broader framework for music composition in
Python called SCAMP. Clockblocks allows for the coordination of parallel and / or nested
clocks running at different tempi, facilitates smooth acceleration and deceleration, and
sleeps precisely while compensating for calculation time. The approach presented here
is compared with other systems for managing musical time, and further development in
terms of coordinating metric phase is considered.
2.1 Introduction
2.1.1 Context
In recent years there has been a proliferation of interest in, and tools designed for,
computer-assisted music composition. Among the options available, one might broadly
distinguish between domain-specific languages, such as SuperCollider or Max/MSP, and
frameworks that operate within general-purpose programming languages, such as abjad
[3] or jMusic [20]. While both approaches have advantages, one major advantage of
31
Clockblocks: A Pure-Python Library for Controlling Musical Time Paper 2
situating a composition framework within a general-purpose language is the wide range

of powerful libraries that are made readily available.
Another important distinction exists between languages and frameworks aimed at the
direct generation of sound, and those aimed at the creation of a score to be played by
live performers. These very different aims necessitate significant differences in design; for
instance, speed and efficiency are critical concerns when generating real-time audio, while
for score-generation they are much less important. On the other hand, traditional music
notation places very significant (and idiosyncratic) constraints on timing and rhythm, as
well as on other musical parameters.
SCAMP (Suite for Computer-Assisted Music in Python) [21] is a GPL3.0-licensed
framework for musical composition that aims to take advantage of the general-purpose
nature and compact, readable syntax of Python, while at the same time situating itself
as a hybrid between sound-oriented and notation-oriented frameworks. Although the
creation of traditionally notated scores is a central aspect of the framework, SCAMP is
nevertheless strongly playback-oriented: rather than interacting with a score, the pro-
grammer interacts with a virtual ensemble, listening to and tweaking the resulting play-
back until he or she is ready to translate the music to traditional western music notation.
The key functions of the SCAMP framework are:
• To provide facilities for flexible and extensible note playback, e.g. via FluidSynth
or over OSC. (Effortless microtonality and glissandi have been built in.)
• To manage the flow of multiple interconnected streams of musical time.
• To record note-events, quantize recorded performances sensibly and flexibly, and

translate the result to legible music notation, either as MusicXML or as LilyPond
(via the abjad library).
32
A key value underlying the development of SCAMP is that of modularity and adher-
ence as much as possible to the Unix Philosophy. For instance, the MusicXML export
capability is available separately as pymusicxml, the flexible musical Envelope class is
available separately as expenvelope, and the system for managing musical time is avail-
able separately as clockblocks. It is this last library that is the subject of this paper.
2.1.2 Goals
Clockblocks arose to address several recurring problems with the scheduling of note
playback events and the recording of note event data in Python:
1. The time.sleep function from the Python Standard Library has limited accuracy,
especially for longer wait times.
2. Playback is slowed down by script execution, noticeably so if extensive calculations

are involved.
3. In multi-part music running in parallel threads, differing calculation times result

in drift between the parts. This is especially problematic if note events are to be
recorded and quantized.
4. A system for controlling and modulating tempo is needed, ideally one allowing for
multiple independent streams operating simultaneously.
Clockblocks solves the first of these problems by defining a sleep_precisely function

which repeatedly sleeps for half the remaining duration until fewer than 500µs are left, at
which point it implements a busy wait for the remaining time. The remaining problems
are addressed through a system of interconnected clocks that are coordinated through
with a single master clock, which does the actual sleep calls. In this way, multiple
33
independent streams of musical time, potentially running simultaneously at different

tempi, remain perfectly synchronized.
2.2 A Simple Example
2.2.1 The Code
The following example will serve to introduce the clockblocks API:
from clockblocks import Clock
clock = Clock(initial_tempo=60)
def log_timing():
print(
"Beat:", clock.beats(),
"Time:", round(clock.time(), 2),
"Tempo:", round(clock.tempo, 2)
)
while clock.beats() < 4:

log_timing()
clock.wait(1)
# change to 120 bpm (2 beats per second)

clock.tempo = 120
log_timing()
clock.wait(1)
# gradually slow to 30 bpm over 8 beats

clock.set_tempo_target(30, 8)
log_timing()
clock.wait(1)
34
The resulting output of this program is:
Beat: 0.0, Time: 0.0, Tempo: 60.0

Beat: 1.0, Time: 1.0, Tempo: 60.0
Beat: 2.0, Time: 2.0, Tempo: 60.0
Beat: 3.0, Time: 3.0, Tempo: 60.0
Beat: 4.0, Time: 4.0, Tempo: 120.0
Beat: 5.0, Time: 4.5, Tempo: 120.0
Beat: 6.0, Time: 5.0, Tempo: 120.0
Beat: 7.0, Time: 5.5, Tempo: 120.0
Beat: 8.0, Time: 6.0, Tempo: 120.0
Beat: 9.0, Time: 6.59, Tempo: 87.27
Beat: 10.0, Time: 7.37, Tempo: 68.57
Beat: 11.0, Time: 8.34, Tempo: 56.47
Beat: 12.0, Time: 9.5, Tempo: 48.0
Beat: 13.0, Time: 10.84, Tempo: 41.74
Beat: 14.0, Time: 12.37, Tempo: 36.92
Beat: 15.0, Time: 14.09, Tempo: 33.1
Beat: 16.0, Time: 16.0, Tempo: 30.0
Beat: 17.0, Time: 18.0, Tempo: 30.0
Beat: 18.0, Time: 20.0, Tempo: 30.0
Beat: 19.0, Time: 22.0, Tempo: 30.0
Note that the faster the tempo, the less time advances for a given beat, and the
slower the tempo, the more time advances. The speed of a clock can be set using any of
three interrelated properties: its rate, its beat length, and its tempo. These are defined
as follows:
R = 1/Lb (2.1)
T = 60 · R = 60/Lb (2.2)
Where Lb represents beat length and is measured in seconds (at least on a top level
35
clock), R represents rate and is measured in in beats per second, and T represents tempo
and is measured in beats per minute. Setting any one of these properties for a clock
automatically sets the other two. In some ways, rate and beat length are the most
natural descriptors, especially when clocks are nested inside of each other. However,
tempo is retained as a property because of its associated musical intuition.
2.2.2 Implementation
Each clock internally uses a TempoEnvelope object to manage changes of tempo.

TempoEnvelope is a subclass of the Envelope class from the SCAMP package expenvelope,
which defines a piecewise exponential curve similar in function to the Env object in
SuperCollider [22]. In practice, this means that any accelerandi or ritardandi can be
given a non-linear shape, increasing the expressive potential of clockblocks.
Example TempoEnvelope
2.5
2.0
1.5
Beat Length
1.0
0.5
0.0
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
Beat
Figure 2.1: Graph of the clock’s TempoEnvelope from the example above
Although the user is likely thinking in terms of rate or tempo, internally the tempo
36
envelope is based on beat length for ease of calculations. Figure 2.1 shows this beat
length curve for the above example; when the tempo jumps to 120 bpm on beat 4, the
beat length cuts to 0.5, and then from beat 8 to beat 16 it slowly increases to 2 during
the ritardando. The TempoEnvelope class keeps track internally of the current beat, and
whenever the user calls clock.wait(beats), the area under the curve is integrated from
the current beat forward to the destination beat to determine the associated wait time
in seconds.
2.3 Parallelism
2.3.1 Parallel Clocks Example
The above example featured a single stream of musical time. However, the true
strength of clockblocks lies in its ability to coordinate multiple parallel streams of time,
as in the following example:
master = Clock()
def child1(my_clock: Clock):

while my_clock.beats() < 6:
print("Child Clock 1 at beat {}".
format(my_clock.beats())
my_clock.wait(1)
def child2(my_clock: Clock):

while my_clock.beats() < 3:
print("Child Clock 2 at beat {}".
format(my_clock.beats())
my_clock.wait(1/3)
37
master.fork(child1)
master.fork(child2)
master.wait_for_children_to_finish()
In this example we create a master clock and define two parallel processes, child1
and child2. The first function prints every beat until beat 6, while the second prints
every third of a beat until beat 3. We then fork these functions on the master clock. The
results are as follows:
Clock 1 at beat 0.0

Clock 2 at beat 0.0
Clock 2 at beat 0.333
Clock 1 at beat 1.0
Clock 2 at beat 1.0
Clock 2 at beat 2.0
Clock 1 at beat 2.0
Clock 1 at beat 3.0
Clock 1 at beat 4.0
Clock 1 at beat 5.0
Note that the child functions take a single argument which gets passed a handle to
the clock being forked. It is also possible to call get_current_clock() to capture the
clock running at any given moment.
2.3.2 Parallel Clocks Implementation
Whenever a child clock calls wait, it registers a wake up time with its parent and then
pauses execution until the parent rouses it. The parent clock maintains a cue of wake up
38
calls from its children, and whenever it calls wait, it looks to see if there is a child clock
with a wake up time in the near future so that it can rouse it at the appropriate time.
An example sequence of events with both parent and child starting at t = 0 might be
as follows:
t = 0:
- Child calls wait(0.5), registers wake up time of
0.5 with parent.
- Parent calls wait(1), sees that a child is set to
be woken up at t = 0.5, and so waits instead for
0.5 beats.
t = 0.5:
- Parent wakes up and rouses child
- Child wakes and calls wait(1.0), registering a
wake up time of 1.5 with parent.
- Parent sees no other child wake events during the
rest of its wait of 1, and so sleeps for the
remaining 0.5.
t = 1.0:
- Parent wakes from its sleep, calls wait(2) this
time, and sees that a child has registered a wake
up time of 1.5. As a result, it waits 0.5 second.
t = 1.5:
- Parent wakes up and rouses child.
- Child wakes up and chooses to terminate process.
- Parent sees no other child wake events during the
rest of its wait of 2, and so sleeps for the
remaining 1.5.
t = 3.0:
- Parent wakes, sees it has no children, suffers
from empty nest syndrome.
39
When a forked function reaches the end of its execution, the child clock associated
with it is terminated. In the example in Sec. 2.3.1, the master clock, acting as the
parent to both child clocks, calls wait_for_children_to_finish, which causes it to
wait indefinitely, rousing its child clocks at the appropriate times until all children have
finished execution.
Note that child clocks can fork their own child clocks, and so on. In this case, a clock
may find itself in the role of both parent and child, waking its children at the appointed
times, and registering a wake up call with its parent whenever it wishes to wait itself.
Only a master clock, a clock with no parent, actually calls time.sleep (or rather, the
more precise version explained in Sec. 2.1.2). All other clocks simply register a wake up
call with their parent when they wish to sleep.
2.4 Compensating for Calculation Time
One of the initial problems that clockblocks was designed to solve was the
fact that, unless compensated for in some way, any calculation time on a thread
will cause that thread to slow down relative to the sum of all of its calls to
time.sleep. It should be clear from the above that this problem is already solved
for all but the master clock, since wake up times are absolute and will not drift. It only
remains to ensure that the master clock itself takes calculation time into account.
When a master clock wakes up, it immediately notes down the current time. Then,
after all relevant calculations have taken place and a new wait call is made, it refers back
to the time that it originally woke up in determining how long to sleep.
In some cases, when calculations are intensive and the wait time is short, it may
already be past the the desired wake up time, and the clock finds itself running behind.
At this point there are two main options:
40
1. Allow the clock to stay behind and handle subsequent wait times as faithfully as
possible.
2. Try to catch up in future calls to wait by not waiting at all until the clock has
caught up.
Both options are available in clockblocks. The first is termed a “relative” timing
policy (since it emphasizes keeping individual wait times as accurate as possible), while
the latter is termed an “absolute” timing policy (since it emphasizes not drifting from the
absolute time at which events should have occurred). If the playback from clockblocks
is being coordinated in any way with that of an external application, an absolute timing
policy would likely be preferred; if not, a relative policy may be preferred.
The default used by clockblocks is actually a third, hybrid approach. This policy
allows for time to be shaved off of subsequent calls to wait, but only to a certain degree.
Thus, rather than catching up all at once, the clock catches up in small increments. In
many cases this is sufficient to remain faithful to the absolute times at which events
should occur without distorting subsequent wait times noticeably.
2.5 Nested Clocks
2.5.1 Tempo Inheritance
The true power of clockblocks as a library lies in the fact that each clock, regardless
of its place in the family hierarchy, is allowed to have and manipulate its own tempo.
However, the actual speed at which a clock runs depends not only on its own tempo, but
also on the tempo of its parent, and its parent’s parent, etc. all the way on up to the
master clock.
41
To understand this better, consider that a clock has two different views on the passage
of time: what beat it is on and how much actual time has passed. As we have seen above,
these two properties are related by the clock’s beat length; time passed is the integral of
beat length with respect to beats passed.
In clockblocks, each clock inherits its sense of time from its parent; a beat in the
parent clock constitutes a “second” of time in the child clock. The word “second” here is
in quotes, because unless the clock in question is the master clock, it is not a true second,
but rather a second as filtered through temporal distortions of its parents.
For instance, in Figure 2.2, we consider three generations of nested clocks: a master
clock running at rate 1/2, its child (e.g. the result of a call to fork) running at rate 3,
and its child’s child, running at rate 1/4. Note that the child’s sense of time is inherited
from the master clock’s beat rate, and the grandchild’s sense of time is inherited from
the child’s beat rate. Thus is should be clear from this picture that the tempi of clocks
in a parent / child relationship multiply. The absolute rate of the grandchild clock – its
rate with respect to wall time – is the product of its own rate, its parent’s rate, and its
parent’s parent’s rate.
MASTER time: 0 1 2 3 4 5 6 7 8
rate = 1/2
absolute rate
tempo = 30
beat: 0 1 2 3 4 = 1/2
beat length = 2
CHILD time: 0 1 2 3 4
rate = 3 absolute rate
tempo = 180 = 1/2 * 3
beat: 0 1 2 3 4 5 6 7 8 9 10 11 12 = 3/2
beat length = 1/3
GRANDCHILD time: 0 1 2 3 4 5 6 7 8 9 10 11 12 absolute rate

rate = 1/4
= 1/2 * 3 * 1/4
tempo = 15
beat length = 4
beat: 0 1 2 3 = 3/8
Figure 2.2: Relationship between the rates of three “generations” of clocks running
at different tempi
42
What is not depicted in the above example is that each clock can in fact be smoothly
changing rate according to manipulations of its tempo envelope. The actual amount of
wall time corresponding to a wait of, say, two beats in the grandchild clock is calculated
by first integrating for two beats under the grandchild clock’s tempo envelope, then taking
the result and integrating for that many beats under the child clock’s tempo envelope,
and then taking that result and integrating for that many beats under the master clock’s
tempo envelope.
2.5.2 A Nested Tempo Example
The following example will server to illustrate how nested clocks can follow different,
but interacting, tempo envelopes. It also introduces several new features for shaping a
clock’s tempo over time:

import math
def child_process_1(clock: Clock):

clock.apply_tempo_function(
lambda t: 60 + 30 * math.sin(t),
duration_units="time"
)
# do something musical here
clock.wait(40, units="time")
def child_process_2(clock: Clock):

clock.apply_tempo_envelope(
[70, 90, 90, 55, 85, 70],
[2, 1.5, 2.5, 1.5, 1.0],
curve_shapes=[0, 3, 0, -2, 0], loop=True
)
# do something musical here
clock.wait(40, units="time")
43
master = Clock()
child_1 = master.fork(child_process_1)
child_2 = master.fork(child_process_2)
master.set_rate_target(3, 15)
master.set_rate_target(1, 40, truncate=False)
master.wait_for_children_to_finish()
Here, we create a master clock and spawn two child processes. One of these pro-
cesses follows a sinusoidal tempo envelope, which we create by calling clock.apply_
tempo_function. (Internally, this tempo function is being approximated by exponential
curve segments, since all tempo envelopes are piece-wise exponential.) The other clock
instead calls apply_tempo_envelope to define this piece-wise exponential curve directly.
Since the loop flag has been set to True, the tempo envelope repeats for as long as the
clock is alive.
The master clock itself changes tempo over the course of the example, going to a rate
of 3 over the course of 15 beats and back to a rate of 1 after 40 beats have passed. Note
that the truncate flag has been set to False in the second call to set_rate_target; by
default when a rate/tempo/beat length target is set, any existing targets are cut off, but
by setting the truncate flag to False, the first target remains in place.
After running the code above, the following lines will generate the plots shown in
Figure 2.3:
# show_plot uses matplotlib under the hood

master.tempo_envelope.show_plot()
child_1.tempo_envelope.show_plot()
child_2.tempo_envelope.show_plot()
child_1.extract_absolute_tempo_envelope().show_plot()
child_2.extract_absolute_tempo_envelope().show_plot()
44
Child 1 Tempo Envelope Child 1 Absolute Tempo Envelope
Master Tempo Envelope
Child 2 Tempo Envelope Child 2 Absolute Tempo Envelope
Figure 2.3: Effect of a master tempo envelope of the absolute tempo envelopes of its
children. Note that, unlike in Figure 1, these graphs are of tempo, rather than the
beat length.
As Figure 2.3 illustrates, the tempo envelope of the master clock affects those of the
two child clocks. The plots in the middle column show the tempo envelopes of the child
clocks with respect to their parent (the master clock), and show the sinusoidal variation
and explicitly defined tempo envelope described above. On the other hand, the plots in
the right column show their absolute tempo envelopes, i.e. their tempos with respect to
wall time, having been altered by the acceleration and deceleration of the master clock.
Before we leave this example, it should be pointed out that the child clocks call wait
with the additional keyword argument units="time". What this does is instruct the
clock to wait however many beats will correspond to 40 units of time, (which is the same
as 40 beats in the parent clock). This affords the ability to coordinate with the parent
clock; for instance, in this case the master clock has been instructed to accelerate and
decelerate over the course of 40 beats, which will be the exact same length as 40 units of
time in the two child processes.
45
Notice also that time units have been specified for the sinusoidal tempo function
defined on the first child clock. This is why, in the graph of its tempo envelope, the
peaks are wider than the troughs; this is a graph of tempo with respect to beats, and it
will take more beats to cover a given amount of time at a faster tempo than at a slower
tempo. On the other hand, the tempo envelope applied to the second child clock is in
the units of beats (which is the default), so the graph appears undistorted.
2.6 Comparison with other approaches
In contrast to audio programming environments like Gibber [23] and ChucK [24],
clockblocks does not concern itself with the audio thread directly, or with sample accu-
racy. As the time management engine of SCAMP, clockblocks is designed for scheduling
events at the rate of notes and sound objects. The actual production of sound samples
happens externally, for instance via FluidSynth or OSC messages to some other external
instrument. Temporal precision is of course desired, but sample accuracy is not necessary.
Clockblocks does bear some similarity to ChucK in its syntax, however: the user
performs operations, sets up processes and then then advances time. As in ChucK,
subprocesses can be forked, and these subprocesses can themselves fork subprocesses.
However, clockblocks makes use of nested tempo relationships in a way that ChucK does
not, or at least not natively.
SuperCollider [2] provides the ability to control multiple independent streams of
tempo via the TempoClock object. In some ways, clockblocks also resembles Super-
Collider in its server/client dichotomy; the client language that is in charge of scheduling
is separate from the process of generating audio samples. However, it is not possible in
Supercollider to nest TempoClocks, and any accelerandi and ritardandi must be accom-
plished through rapid incremental changes of tempo.
46
Thus, the main contribution of clockblocks is that it combines the nested structure
of an environment like ChucK with the ability to smoothly manipulate tempo at any
layer of this structure. It also provides this functionality in Python, a general purpose
programming language that offers access to a vast array of packages from a wide variety
of disciplines, and one that has at present a dirth of options for managing musical time.
2.7 Directions for Further Development
Although in its current implementation clockblocks does not offer sample accuracy (nor
is this necessary for its role within SCAMP), the nested tempo approach presented here
has the potential for broader application, including some contexts (like the scheduling
of microsonic events) where sample-accuracy would be desired. Therefore, one natural
direction for further work would be to translate this system to a language like C++,
using it generating a cue of sample-clock timestamped events. Even within its current
Python implementation, one planned development is to allow time-sensitive playback
events, such as OSC messages or calls to an external synthesizer, to be scheduled on a
queue for later dispatching by an audio callback (for instance, via a PyAudio’s PortAudio
bindings).
Another area of clockblocks currently being developed is the ability to specify and
synchronize rhythmic phase. Here we take inspiration from “Tempocurver” [25], devel-
oped by Matthew Wright in collaboration with composer Edmund Campion, as well as
from CNMAT’s subsequently developed Max / MSP external, “Timewarp” [26].
As we have seen, clockblocks makes it possible to coordinate clocks so that they
reach specific tempi at specifically appointed times. However, it may also be musically
important to coordinate rhythmic phase; for instance, we may want two clocks that are
following different tempo curves to land on the beat at the exact same moment in time.
47
This requires fine tuning of the relationship between beats passed and time passed,
maximum time passed = 2 beats * 2 s / beat = 4 s
minimum time passed = 2 beats * 0.75 s / beat = 1.5 s
Time Passed
Figure 2.4: Illustration of the relationship between curvature and the length (in time)
of an accelerando.
As Figure 2.4 illustrates, the relationship between the number of beats passed and
the amount of time passed during a segment of a TempoEnvelope is mediated not only
by the start and end tempo, but also by the curve shape. In the illustrated accelerando,
by varying the curvature, we can adjust the amount of time passed to anywhere between
1.5 and 4 seconds. Thus, if we wished for this 2 beat accelerando to last precisely 3
seconds, we would need only to find the appropriate curve shape.
By using adjustments of this nature, it should be possible to specify the desired metric
phase (in terms of time, or beats in the parent clock) at the end of an accelerando or
ritardando.
48
2.8 Conclusions
Reviewing the initial goals of clockblocks, it is hopefully now clear to the reader that
the system described:
1. Is capable of waiting significantly more precisely than the standard call to

time.sleep.
2. Takes efforts to compensate for calculation time, and has an intelligent system for
adjusting when calculation time lasts longer than an intended wait time.
3. Keeps track of multiple interconnected threads of musical time and ensures that
these threads remain in lockstep with one another.
4. Allows for complex control and modulation of tempo, and creates the potential for
nested tempo relationships.
These properties make clockblocks an excellent foundation for a note-event-based play-

back and recording framework in Python. Several different streams of time can easily
coexist, and the resulting music can be recorded and quantized in relation to any one
of these streams. This has particularly exciting implications for the playback and score
generation of polytemporal music.
49
Bibliography
[1] M. Puckette, Pure data: another integrated computer music environment, in in

Proceedings, International Computer Music Conference, pp. 37–41, 1996.
[2] J. McCartney, Rethinking the Computer Music Language: SuperCollider, Computer

Music Journal 26 (Dec., 2002) 61–68.
[3] T. Bača, J. Oberholtzer, J. Treviño, and V. Adán, Abjad: An Open-Source

Software System for Formalized Score Control, in Proceedings of the First
International Conference on Technologies for Music Notation and Representation –
TENOR’15, 2015.
[4] J. Bresson, C. Agon, and G. Assayag, OpenMusic: visual programming

environment for music composition, analysis and research, in Proceedings of the
19th ACM international conference on Multimedia - MM ’11, (Scottsdale, Arizona,
USA), p. 743, ACM Press, 2011.
[5] M. S. Cuthbert and C. Ariza, music21: A Toolkit for Computer-Aided Musicology

and Symbolic Music Data, in Proceedings of the 11th International Society for
Music Information Retrieval Conference (ISMIR 2010), International Society for
Music Information Retrieval, Aug., 2010.
[6] M. Good, MusicXML: An Internet-Friendly Format for Sheet Music, in

Proceedings of XML, (Boston, Massachusetts, USA), 2001.
[7] H.-W. Nienhuys and J. Nieuwenhuizen, LilyPond, a System for Automated Music
Engraving, Proceedings of the XIV Colloquium on Musical Informatics (2003).
[8] D. Rossum, “The SoundFont

R 2.0 File Format.”
[9] N. Whitehead, “pyfluidsynth.” https://github.com/nwhitehead/pyfluidsynth,

2018.
[10] A. Freed and A. Schmeder, Features and Future of Open Sound Control version
1.1, in Proceedings of the 2009 New Interfaces for Musical Expression (NIME)
Conference, 2009.
50
[11] O. Jolly, “sf2utils.” https://gitlab.com/zeograd/sf2utils, 2018.
[12] C. Arndt, “python-rtmidi.” https://github.com/SpotlightKid/python-rtmidi,
2019.
[13] attwad (username), “python-osc.” https://github.com/attwad/python-osc,
2018.
[14] M. Evanstein, Clockblocks: A Pure-Python Library for Controlling Musical Time,
in Proceedings of the 2019 International Computer Music Conference, (New York,
New York, USA), 2019.
[15] S. W. Smoliar, A Parallel Processing Model of Musical Structures, .
[16] C. Barlow, On Musiquantics: Von der Musiquantenlehre translated. Royal
Conservatory The Hague, 2012.
[17] P. Nauert, A Theory of Complexity to Constrain the Approximation of Arbitrary
Sequences of Timepoints, Perspectives of New Music 32 (1994), no. 2 226–263.
[18] “The Linux Sampler Project.”
[19] “The SFZ Format.”
[20] A. Brown, Making Music with Java. Andrew R. Brown, May, 2009.
[21] M. Evanstein, “Scamp: a suite for computer-assisted music in python.”
https://github.com/MarcTheSpark/scamp, 2018.
[22] SuperCollider 3 Documentation Contributors, “Env — supercollider 3.10.0 help.”
http://doc.sccode.org/Classes/Env.html, 2018.
[23] C. Roberts, M. Wright, J. Kuchera-Morin, and T. Höllerer, Gibber: Abstractions
for Creative Multimedia Programming, in Proceedings of the ACM International
Conference on Multimedia - MM ’14, (Orlando, Florida, USA), pp. 67–76, ACM
Press, 2014.
[24] G. Wang, P. R. Cook, and S. Salazar, ChucK: A Strongly Timed Computer Music
Language, Computer Music Journal 39 (Dec., 2015) 10–29.
[25] M. Wright and E. Campion, “Tempocurver.”
https://github.com/CNMAT/CNMAT-Externs/tree/master/java/tempocurver,
2001.
[26] J. MacCallum and A. Schmeder, Timewarp: A Graphical Tool For The Control Of
Polyphonic Smoothly Varying Tempos, in Proceedings of the 2010 International
Computer Music Conference, ICMC 2010, New York, USA, 2010, 2010.
51

Evanstein MAT Thesis SCAMP

Uploaded by

Copyright:

Available Formats

Evanstein MAT Thesis SCAMP

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Evanstein MAT Thesis SCAMP

Uploaded by

Copyright:

Available Formats

University of California

A project document submitted in partial satisfaction

Marc Paul Evanstein

Professor Curtis Roads, Chair

Marc Paul Evanstein

Marc Paul Evanstein

2 Clockblocks: A Pure-Python Library for Controlling Musical Time 31

This paper introduces SCAMP, a computer-assisted composition framework in Python

Computer-assisted composition offerings can be broadly divided into two groups:

• Playback using an external synthesizer via an outgoing MIDI stream. (SCAMP

• Playback using an external synthesizer via OSC [10]

• User-defined playback implementations created by implementing atomic playback

1.1.3 Designed for Flexibility

• Coordinate multiple time streams • Note based (though broadly)

• Easily understood syntax, with sig-

Table 1.1: Key features and limitations of SCAMP

• Easier maintenance: keeping different aspects of functionality separate keeps the

LilyPond output pythonosc

Figure 1.1: Internal and external dependencies of SCAMP

1.2 Introductory Examples

1.2.1 ”Hello World”

# import the scamp namespace

# construct a session object

# add a new violin part to the session

# looping through the MIDI pitches

Figure 1.2: The combined functionality of a Session object

# start playing a middle C with volume 0.7

It is important to note that durations do not have to be of rational lengths. For

# import the random module from

while True: # loop forever

1.2.3 Generating Notation

from scamp import *

# begin transcribing (defaults to using

for pitch in [60, 64, 67, 72]:

# stop transcribing and save the

# quantize and convert the performance to

Figure 1.3: Examples of notation generated from introductory examples. Snippets

# ---- setup code omitted for brevity ----

for _ in range(2): # loop twice

This results in Fig. 1.3d. Considerable customizability of QuantizationScheme is

Figure 1.4: Illustration of the quantization process.

1.3 Details and Further Possibilities

1.3.1 Multi-Part Music

from scamp import *

# one way of setting an initial tempo

# define a function for the oboe part

# define a function for the bassoon part

# have the session wait for the child

1.3.2 Multi-Tempo Music

One of the most powerful aspects of SCAMP’s approach to time-management is that

from scamp import *

# Have the session as a whole speed up to

# here we define a trumpet part to run as

# Fork the trumpet part as a child process

# Play quarter notes for 12 beats

# Stop transcribing and show the result

a.) The Rubytles

An important aspect of notating polytempo music in SCAMP is that performances

# fork returns the Clock associated