Transcript kolesnik05melodic

Melodic Similarity
MUMT 611, March 2005
Assignment 4
Paul Kolesnik
1
Conceptual and Representational
Issues in Melodic Comparison

(Selfridge-Field)
 Melody
 Melodic material can be:
 Compound, self-accompanying, submerged, roving,
distributed
 Theme
 A shorter sample from longer melodic materials that can be isolated
and classified
2
Conceptual and Representational
Issues in Melodic Comparison

Components of Melodic Representation
 Representative

pitch, duration
 Derivable

intervallic motion, accents
 Non-derivable

articulation, dynamics
3
Conceptual and Representational
Issues in Melodic Comparison

Pitch Processing
 Different

ways of pitch labeling
Base 7, base 12, base 21, base 40
 Approaches
to melodic pitch representation
profiles of pitch direction (up-down-repeat)
 pitch contours, melodic contours (sonographic data,
shapes of melodies)
 pitch-event strings (employ base-system representation)
 intervallic contours (intervallic profiles)

4
Conceptual and Representational
Issues in Melodic Comparison

Multi-dimensional data comparison
 Models:

Kernel-filling model






melody seems to evolve from a kernel consisting of outer note of a
phrase
uses both pitch and metrical data
Accented-Note Models
Coupling Procedures
Synthetic Data Models
Parallel processing models
5
A geometrical algorithm for
melodic difference (Maidin)
 identifies similar
1, 8-bar segments in irish folk-dance
music
 based on:





juxtapositioning of notes in two melodic segments
pitch differences (using base-12 or base-7)
note durations
metrical stress
transpositions (trying different transpositions and taking the
minimum differenc value)
6
String-matching techniques for
musical similarity and melodic
recognition


(Crawford, Iliopoulos, Raman)
Describes string-pattern matching algorithms
 approaches
with known solutions
 approaches with unknown solutions



Notion of themes, motifs
Notion of characteristic signature
Motifs have melodic similarity if they have matching
signatures
7
String-matching techniques for
musical similarity and melodic
recognition



String: sequence of symbols drawn from alphabet
Uses two-dimensional mode: pitch, duration
Pattern matches:
 exact
(pitch info is matched)
 transposed (intervallic info is matched

special case of transposed: octave-displaced match
8
String-matching techniques for
musical similarity and melodic
recognition

Exact-match algorithms









Exact matching
Matching with deletions (no duration patterns preserved)
Repetition identification (non-overlapping patterns in different
voices/same voice)
Overlapping repetition identification
Transformed matching (retrograde, inversion)
Distributed matching (across voices)
Chord recognition
Approximate matching (Hamming distance)
Evolution detection
9
String-matching techniques for
musical similarity and melodic
recognition

Inexact-match:
Unstructured exact matching (find a pattern in voiceunspecified mixture of notes)
 Unstructured repetitions (identified repeating patterns
that may/may not overlap)
 Unstructured approximate matching

10
Sequence-based melodic comparison:
a dynamic programming approach





(Smith, McNab, Witten)
Describes dynamic programming (string matching)
algorithm
Used on database of 9400 folk songs
Based on edit distance (cost of changing string a into
string b) using edit operators: replacement, insertion
and deletion
General:can be applied to any type of string (pitch,
rhythm for music)
11
Sequence-based melodic comparison:
a dynamic programming approach



Cost/weight assigned to each operation, based on the
input string components
Uses local score matrix (scores for each element of
the two strings), global score matrix (score of a
complete match between two strings)
Techniques of fragmentation/consolidation
 Eg.
four notes can match one longer note and vice versa.
12
Signatures and Earmarks: Computer
recognition of patterns in music

(Cope)

Creating new scores based on originals using
‘Experiments in Musical Intelligence’ (EMI) system
Musical signature

a
motif common to two or more works of a given
composer, 2-5 beats in length and composites of melodic,
rhythmic, harmonic components

Uses base-12 system, a number of controllers
13
Signatures and Earmarks: Computer
recognition of patterns in music

Earmarks
 More
generalized than signatures, refer to identifying
specific locations in the structure of a musical score (what
movement of a work we are hearing)
 Eg. trill followed by a scale, upward second followed by a
downward third
 Distinguishing quality: location
14
A Multi-scale Neural-Network Model
for Learning and Reproducing Chorale
Variations

(Hornel)
 Style
is learned from musical pieces of baroque
composers (Bach, Pachelbel), new pieces are
produced
 System able to learn and reproduce higher-order
elements of harmonic, motivic and phrase structure
15
A Multi-scale Neural-Network Model
for Learning and Reproducing Chorale
Variations
 Learning
is done using two mutually interacting NN,
operating on different time scales, unsupervised learning
algorithm to classify and recognize structural elements
 Complementary intervallic encoding
 Given a chorale melody, a chorale harmonization of the
melody is invented, and one of the voices of harmonization
is selected and provided with melodic variations
16
Judgments of Human and Machine
Authorship in Real and Artificial
Folksongs

(Dahlig, Schaffrath)
 Listeners presented
with series of original and artificially
created folksongs
 Perception of the nature of composition varied with
perception of the music itself
 Associations with original: rhythmic similarity of phrases,
final cadence on the 1st degree, intermediate phrase
beginning that did not start on the 1st degree.
17
MELDEX: A Web-based Melodic
Locator Service

(Bainbridge)
 Query by humming
 Four databases: North-American/British, German, Chinese,
Irish folksongs; 9400 melodies
 Two alternative algorithms:


simple, fast, state matching algorithm
slower, sophisticated dynamic programming algorithm
18
Themefinder: A Web-based Melodic
Search Tool

(Kornstadt)
 Database of 2000 monophonic theme representations for
instrumental works from 18th-19th centuries

Search parameters
 pitch
direction (gross contour or refined contour)
 letter name of pitch
 pitch class
 intervallic name
 intervallic size
 scale degree
19
A Probabilistic Model of Melodic
Similarity
 Hu,
Dannenberg, Lewis (2002)
Compares dynamic programming to probabilistic
approach in sequence matching
 Used query by humming as input
 Collected and processed 598 popular song files
 Processing done using MUSART thematic
extractor (10 themes per song), 5980 entries with
average 22 notes per song

20
A Probabilistic Model of Melodic
Similarity

Dynamic Programming Algorithms
Edit Distance
 Frame-based (pitch contour) matching


Probabilistic Approach


Probabilistic Distribution Histogram
Results

Probabilistic model outperformed dynamic programming
algorithms by a narrow margin
21
Name That Tune: A Pilot Study
in Finding a Melody From a
Sung Query






(Pardo, Shifrin, Birmingham)
A query by humming system
Two-dimensional: pitch and rhythm
Comparison between string-alignment (edit cost) dynamic
programming and HMM algorithms (each theme represented
as a model)
Also compared to human performance
Results


String-alignment algorithms slightly outperform HMM
Human performance is superior to both HMM and string
algorithms
22
Melodic Similarity - Providing a
Cognitive Groundwork

(Hoffman-Engl, 1998-2004)
Original algorithms: string comparison-based
 New: geometric measure, transportation distances,
musical artist similarity, probabillistic similarity,
statistical similarity measures, transformational models,
transition matrices.
 Comparison problem: validity of results

23
Melodic Similarity - Providing a
Cognitive Groundwork
 Dynamic
values as a separate dimension
 Similarity must not be based on physical but on
psychological dimensions
 Meloton, Chronoton, Dynamon
 Generalizations
 Larger the transposition interval, smaller similarity
 Larger tempo difference, smaller similarity
24
Melodic Similarity - Providing a
Cognitive Groundwork
 Factors contributing to melodic similarity
 Melotonic distance (pitch value difference)
 Melotonic interval distance (distance between pitch intervals)
 Chrontonic distance (difference between durations)
 Tempo distance
 Dynamic distance (difference between dynamic values)
 Dynamic interval distance (between relative dynamic values)
 A cognitive model
based on those factors is presented
25
Conclusion

HTML Bibliography
http://www.music.mcgill.ca/~pkoles

Questions
26