|
Remember the
wonderful liner notes on the back of Gould's album of the Liszt piano
transcription of the Beethoven symphony? The Dakota psychiatrist accused GG of
megalomania for wanting to be an entire symphony orchestra. And the Socialist
reviewer criticized him for stealing the bread from the mouths of 60 musicians
and their families.
On "Saturday Night Live," the
talentless lounge singer Bill Murray used to point to his cheesy, annoying
little percussion machine and ask the audience to give a big round of applause
to "the Univox 4000."
We all laughed that a box might ever
replace a human musician (even a drummer, a stretch both for "human"
and "musician").
Let laughter cease. I present, without
further comment, The Future:
==================
Amherst College
(Amherst, Massachusetts
USA)
Mathematics and Computer Science
Colloquium
Professor Chris Raphael
University of Massachusetts, Amherst
[USA]
Music Plus One
I discuss my ongoing work in creating a
computer system that plays the role of a sensitive musical accompanist in a
non-improvisatory composition for soloist and accompaniment.
An accompanist must synthesize a number
of different sources of information. First of all, the accompanist must perform
a real-time analysis of the soloist's acoustic signal, enabling the accompanist
to "hear" the soloist. The accompaniment must also understand the
basic template for musical performance that is described in the musical score
(notes, rhythms, etc.), thereby allowing the system to "sight-read"
(perform with no training) credibly. However, the acocompanist must also be able
to improve over succcessive rehearsals, much as live musicians do; thus the
accompanist must be capable of learning from training data.
I present a probabilistic model -- a
Bayesian Belief Network that represents these disparate knowledge sources in a
coherent framework. Nodes in the network represent observable variables, such as
estimated note onset times, and unobserable variables, such as local tempo and
rhythmic stress. The connectivity of the graph expresses various conditional
independence assumptions which are key in making the computations feasible in
real-time.
In a series of rehearseas the model is
trained from both solo and accompaniment data to represent a rhythmic
interpretation for a specific piece of music. During live performance, the
accompanist "listens" to the soloist by using a hidden Markov model
and makes principled real-time decisions that incorporate all currently
available information. I will provide a live demonstration of my system on
several examples including Robert Schumann's 1st Romance for Oboe and
Piano.
Wednesday 27 March 2002, 4 p.m.
Seeley Mudd 207
Refeshments will be served in Seeley
Mudd 208 at 3:30 p.m.
[NOTE: I don't know if Raphael is
the oboe or the piano.]
|