From NY Times:
August 28, 2003
By ANNE EISENBERG
SONY has a hit product in Aibo, its robotic pooch that can
ride a skateboard, respond to voice commands, take
snapshots with its digital camera and play with a ball.
Now, for those who want a more refined electronic sidekick,
a French researcher at Sony's Computer Science Laboratory
in Paris has come up with the makings of a musical
companion rather than a canine one.
François Pachet, a skilled pianist and jazz guitarist who
is also a scientist at Sony, has designed an electronic
accompanist that can riff with a musician, seamlessly
extending and improvising on the human player's musical
This talented accompanist does not come in a winsome
plastic skin like Aibo's; at this point it is a software
prototype with no more personality than a player piano. But
it makes convincing music, its admirers say, by rapidly
analyzing what a musician is playing and then joining in
with music of a similar style.
The process of creating this new music is entirely digital.
Dr. Pachet's software program dissects the music, which
must be played on an instrument capable of sending a stream
of notes to the processor in MIDI, a digital standard for
devices like synthesizers that control the emission of
music. When the performer pauses, the computer program
instantly begins generating new music with a MIDI-equipped
The computer-generated music picks up so quickly and
smoothly that many people who have heard demonstrations say
it is impossible to tell when the human musician stops and
the program begins.
Programs like Dr. Pachet's, which is called the
Continuator, may some day be part of the array of skills
offered by entertainment robots that people buy for
"Robots like Aibo are designed to be interactive in an
entertaining way," Dr. Pachet said. "And part of the fun
comes from their being able to learn." The Continuator
learns, he said, because the more data it has, the more
convincingly it analyzes and extends a particular musical
Andrew Schloss, a professor of computer music at the
University of Victoria, heard Dr. Pachet demonstrate his
system on a Yamaha Disklavier, a keyboard that can play
automatically by reading electronic files - the keys move
up and down like those on a player piano, without benefit
of a human being. Because the keyboard is also acoustic, it
can be played the standard way by a musician. Dr. Pachet
used both modes in his demonstration, playing on his own,
then letting the computer pick up, then returning as the
mood took him.
"It was the best improvising program I've ever heard," Dr.
Schloss said of the demonstration. He distinguished no
telltale break between the two performers that would signal
that one was real, the other virtual.
Dr. Schloss suggested that the system could be particularly
useful for musicians who want to analyze their own styles.
"You can learn a tremendous amount from using it for an
hour or two," he said. "You are listening to yourself mixed
up, turned around and modified. It's one hundred percent
based on what you've played."
Automatically generating music in a certain style is not a
new idea. In earlier days, people cast dice repeatedly and
used the numbers to pick musical sequences that together
formed a dance tune, thus creating music without relying on
the rules of composition. "Musikalisches Wurfelspiel," or
"A Musical Dice Game'' (1787), for instance, attributed to
Mozart, involves composing a minuet by rolling dice to
combine prewritten measures of music.
Nowadays music can be generated automatically, with a
mathematical model, for example, that calculates the
probability that certain notes might follow a particular
input. Relying on this model, Dr. Pachet wrote programs
that continuously divide the note stream of the musicians
into phrases. Each phrase is sent to the analyzer and
processed. As the musician plays, the system generates a
continuation from the database. The program can also keep
up with changes in rhythm and chords, so that it can, say,
produce not only a continuation in the same style as the
guitarist but also harmonize with the pianist.
One of the main calculations that concerned Dr. Pachet was
the pause that exists before the system continues a
musician's phrases. It had to be virtually undetectable for
performances to be convincing. To decide on the interval,
he listened to many pieces, among them music played by the
jazz guitarist John McLaughlin, who has a reputation for
speed and who, it turned out, plays a note every 60
milliseconds on average. Using that and other times as a
basis, he and his team programmed the Continuator to learn
and produce sequences in less than 30 milliseconds. Using a
Java prototype of the Continuator running on a Pentium III
laptop, they have since gotten it to produce continuations
in less than 5 milliseconds, Dr. Pachet said.
But could people distinguish between the human player and
the virtual one in this musical version of a Turing Test?
The answer in most cases is no. So far, Dr. Pachet said,
very few people can tell whether a human or a computer is
at work, especially when rapid jazz is played. "We can tune
the continuation so that it is virtually indistinguishable
from the human input," he said.
Last month the Continuator made an appearance at the annual
Siggraph annual conference in San Diego, sponsored by the
Association for Computing Machinery. Mary Farbood, a
classically trained pianist and graduate student at the
M.I.T. Media Lab, tried it out there. Her research is in a
related area: she helped design a computer program,
Hyperscore, that enables children to compose music with
colors and graphic elements. She found the Continuator
innovative. "It takes a novel approach to computer-human
improvisation," she said, adding that it is flexible and
intuitive to work with.
However elegant the Continuator may sound, Dr. Pachet said,
its performance comes down to design rather than
"There is no magic in it," he said. "Mainly it is designing
software that will be able to do all these computations