Information Technology Reference
In-Depth Information
repertoire. But the use of random generation, intrinsic to evolutionary algorithms,
here also, gives results of varying quality, necessitating manual editing (the author
describes the results as 'partially correct'). Note that manual editing echoes the ap-
proach of Harold Cohen with his AARON panting program (McCorduck
1991
), as
well as that of David Cope (
1996
), who use partial manual editing to finish their
compositions. We will see below how we substitute manual intervention by
inten-
tional controls
, and the implication on our models.
Probabilistic grammars were used by Keller and Morrison (
2007
) to generate jazz
improvisation. The outputs of their system 'compare favourably with those played
by college-level jazz students of at least an intermediate playing level, if not better'.
The grammar rules are manually encoded, and based on an explicit representation
of note harmonic status (chord tone, passing tones, etc.). Note-level information
is required when teaching improvisation, but we do not think they are necessary
for generating improvisation. As we will see, we adopt an approach in which this
information is not represented explicitly. However, our generated phrases do contain
a natural blend of, e.g. chord tones and passing notes. Note-level characteristics
naturally emerge from the generator, rather than being prescribed by the system.
Franklin (
2006
) showed that recurring neural networks could learn entire songs
given a melody and the associated chord sequence, and produce new improvisations
on these chord sequences. This system demonstrates that non-symbolic approaches
can capture some of the knowledge of jazz musicians, but the results shown are also
college-level.
Side-slipping is briefly mentioned as a possible composition operation in the
Im-
proVizor
system (Keller et al.
2005
), but the process and more generally the devices
for playing out-of-key 'the right way' have not yet been the subject of modelling in
improvisation generation studies. Finally, it can be noted that commercially avail-
able software like
Band-in-a-box
(PG Music), or the
Korg Karma
™ family of syn-
thesisers produce reasonable improvisations in real time, over arbitrary harmonic
constraints, using algorithmic approaches. These systems may produce musically
interesting outputs, but their analysis is difficult because of a lack of published tech-
nical information.
5.3.2 Markov Chain Approaches
Other approaches to jazz improvisation based on Markov chains have been explored
recently, showing notable success. These systems follow a long tradition in com-
puter music modelling, dating back to the works of Shannon on information theory
(Hiller and Isaacson
1958
, Brooks et al.
1957
). Markov processes are based on the
'Markov hypothesis' which states that the future state of a sequence depends only
on the last state, i.e.:
p(s
i
|
s
1
,...,s
i
−
1
)
=
p(s
i
|
s
i
−
1
)
(5.1)
