Transcript Powerpoint

I548 Presentation
Applications Statistical Graphical
Models in Music Informatics
Yushen Han
Feb 10 2011
Statistical Graphical Models
• graph-based representation for a probabilistic
distribution in high-dimensional space while
specifying conditional independence structure
– directed acyclic graph(DAG) - Bayesian Network
– undirected graph(UG) - Markov Random Field
– Mixed graph
Saturated (Undirected) Graph
Markov Condition on a Bayesian Network
probabilistic
distribution
Highdimensional
space
Conditional
From www.eecs.berkeley.edu/ewainwrig/
independence
Markov Condition: XA and XB are conditionally independent given XS whenever S separates A and B
Musical Application: Bayesian
Identification of Chord from Audio
key
mode
octave root
Attributes
tuning
Observation:
Peak frequencies and
corresponding amplitude
factorization
of the joint
by Bayesian
Inference
Chroma
(pitchclass)
Application I: Score Following/Alignment
Music score
(given)
Performance
audio of this
piece (given)
Establishing a
correspondence
between the two
above
Score Following:
Online (real time)
Score Alignment:
Offline (no need to
be in real time)
Application I: Score Following
Best “guess” of the current location given what the computer
had heard SINCE THE BEGINNING UP TO THE MOMENT
Observation - “Frames” of Audio
Audio
(waveform)
Spectrogram
(via Short-time
Fourier
transform)
Score Following
• Difficulties
–
–
–
–
–
–
Tempo rubato (expressive and rhythmic freedom)
Pitch / amplitude vibrato ( )
Polyphony music
Noise
(occasional) wrong notes etc.
Realtime computational requirement
• Solutions
–
–
–
–
Assuming tempo change is smooth (mostly desirable)
Robust probabilistic data model on normalized semigram
Training to learn a prior (e.g. note length distribution)
Optimized particle filtering for 2-D State-space model
Proposed Solution with 2-D Statespace Model (Bayesian Network)
• Assuming smooth tempo change for tempo rubato
• Two-dimensional state-space model
• Proposing a unit of tempi:
• S(t) - Musical time elapse per audio frame at frame t
• Interpretation: during one audio frame of fixed length
(roughly 64ms, 512 samples at 8000Hz sampling rate), how
much musical time (in terms of 1/384 notes) is elapsing
• In another word, how much of the score the performer
covers every 64ms (not precisely the conventional tempo)
2-D State-space Model
Current
“speed” at this
frame
Accumulative
“speed” up to
this frame –
location
0.64ms
0.64ms
State variables
in one audio
frame (notice
conditional
independence)
Physical Analogy
– Integration of the Speed
speed
location
1
1
1
1
1
2
3
4
Physical Analogy
– Integration of the Speed
with stochastic components
speed
1.0
1.1
1.05
0.95
Speed
fluctuation
location
1.0
2.12
Observation
error also
involved
3.19
4.11
Physical Analogy
- to make a tractable problem
speed
1.0
Assuming discrete state transition with
{-0.1, -0.05, 0, 0.05, 0.1} with prob
distribution {0.1, 0.2, 0.4 ,0.2, 0.1} for
smoothness.
1.1
1.05
0.95
Speed
fluctuation
location
1.0
2.12
Observation
error also
involved
3.19
Assuming Gaussian noise in
observation
4.11
Relationship to Kalman Filter
• Particle filtering
• (Using White board)
Score Following – Data Model
• Data model using semigram
• Regarding discretized observation as a
histogram
Chord template (pre-learned)
Score Following – Demo in R
• Data model in R
• Visualization of results in XCode
Application II:
Graph Model to Estimate Expert
Pianists’ Perceptual Present
- with the help of audio-score
alignment technique
Background
• Curtain eras of classical music – no
improvisation, no wrong notes etc.
• For a certain piece of music, performance
varies in tempo, dynamic, articulation, vibrato
etc. , depending on the interpretation of the
performer
• (This research) focuses on the tempo change
of piano music
Chopin Mazurka Op. 30, No.2
Rubinstein
Horowitz
Michelangeli
http://www.youtube.com/watch?v=PjYV7lJezvc
http://www.youtube.com/watch?v=vGAQONeLnXk
http://www.youtube.com/watch?v=qJmaz1OEGTU
Motivation
• Musical Perceptual Present
– Recent studies in diverse fields of inquiry,
including music philosophy and psychology, lend
converging evidence that musical attention of
both performers and listeners is primarily focused
successively small “chunks” of material
(hypothetically 2–10 seconds in the past) rather
than larger formal relationships.
Motivation
• Instead of individual style, we are in search of
a “common interpretation” shared among a
collection of expert pianists
• Focus purely on tempo change per beat (since
the attack of piano note is easy to capture).
Data
• Human corrected accumulative time per beat
which is equivalent to IBI Inter Beat Interval
• N = 32 performances ( include different
performances of the same pianist )
• For the existence of the MLE, we proceed a
small chunk of data at a time I = { 7, 8, 9, 10 }
Data cont. - preprocessing
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
# !!!performance-id: pid9062-19
# !!!title:
Mazurka in B minor, Op. 30, No.
2
# !!!trials:
1
# !!!date:
2007/02/15/
# !!!reverse-conductor: Craig Stuart Sapp
# !!!performer:
Idil Biret
# !!!performance-date: 1990
# !!!label:
Naxos 8.550359
# !!!label-title:
Chopin: Mazurkas (Complete)
# !!!offset:
0
0.578
0:3
1.398
1:1
1.708
1:2
2.228
1:3
2.668
2:1
3.088
2:2
3.748
2:3
4.336
3:1
4.588
3:2
4.998
3:3
5.498
4:1
5.828
4:2
6.428
4:3
7.028
5:1
7.355
5:2
7.968
5:3
8.428
6:1
8.838
6:2
9.418
6:3
Original Data:
Accumulative time
per beat
•
•
•
•
•
•
•
•
•
•
•
•
•
0.720
0.855
0.745
0.800
0.490
0.610
0.530
0.540
0.550
0.540
0.530
0.570
…
InterBeatInterval
(IBI)
•0.135
•-0.110
•0.055
•-0.310
•0.120
•-0.080
•0.010
•0.010
•-0.010
•…
IBI difference
between beats
Data cont. - preprocessing
• “Normalization”
– Since the overall duration of each performance
varies significantly E.g. Mazurka Op. 30. No. 2
Sec.
Performance index
We “stretch” the overall duration of each performance to line
up with the median of all performances - can be problematic
Model
• For each “trunk” of timing data X of I dimensions
(beats) across N performances:
• N performances are considered i.i.d. repetitions
• we assume: X ~ Ν( 0, Σ )
• where the difference in IBI equals to 0 suggests
that the tempo is nearly constant “on average” (of
course, but could be problematic)
• We study the structure of I by I covariance matrix Σ
• Can obtain an estimate of
Model cont. – check the normality
• See the movie in R
Graph Models
• A toy example for I = 4 case -> 3 hypotheses
– H: Fully saturated model (I-1=3 order Markov chain)
diff.
IBI
– H0: A smaller model (I-2=2 order Markov chain)
– H00: The smallest model (I-3=1 order Markov chain)
Graph Models cont.
• What does this graph mean?
• Conditional Independence!
Graph Models cont.
• Conditional independence in the graph
suggests different structures in the covariance
matrix
?
?
?
To apply reconstruction
algorithm
Graph Models cont. - Testing
• Testing each pair of hypotheses
-2Log(Q) ~
• Accepting the result only when every single pair of
hypotheses of the smallest difference between the
alternative and the null hypotheses are not rejected (as
small step as possible)
• Apply an appropriate degree of freedom ( = difference
in number of edges between 2 graphs )
Results - Testing
• See R plot
Results – Interpretation
• “smoothed” results by using a sliding window
of different lengths
• A “voting” mechanism
• Room to interpret …
Results – Interpretation
“smoothed” results by using a sliding window of different lengths
Results – Interpretation
“smoothed” results by using a sliding window of different lengths
Results – Interpretation
Using “anchor points” to summarize the results