Transcript - Lorentz Center

Music Similarity
Alan Marsden
Lancaster Institute for the Contemporary Arts
Lancaster University
Lorentz workshop on music similarity, Leiden, 23 January 2015
Music Similarity
Alan Marsden
Lancaster Institute for the Contemporary Arts
Lancaster University
Lorentz workshop on music similarity, Leiden, 23 January 2015
Music Ismirality
Alan Marsden
Lancaster Institute for the Contemporary Arts
Lancaster University
Ismirality: talking about music in a manner typical of ISMIR papers
Lorentz workshop on music similarity, Leiden, 23 January 2015
Aim
• To avert the dangers of ismirality
– Music is not just a document
– Music is not just acoustic data
– Music is a trace of human behaviour
• Summing up and personal reactions to the workshop
–
–
–
–
–
–
a principle
a conjecture
a proposal
an observation
a challenge
a reconsideration
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 1: What’s it all about?
• Similarity or categorisation?
– Or are they the same?
• Identity and degrees of difference
– Distance
– Metrics
– Dimensions
Lorentz workshop on music similarity, Leiden, 23 January 2015
My response
• How the workshop looked set to go ...
Lorentz workshop on music similarity, Leiden, 23 January 2015
My response
• How the workshop looked set to go ...
???
Lorentz workshop on music similarity, Leiden, 23 January 2015
?
My response
• How the workshop looked set to go ...
???
Lorentz workshop on music similarity, Leiden, 23 January 2015
?
My response
• How the workshop looked set to go ...
???
Lorentz workshop on music similarity, Leiden, 23 January 2015
?
My response
• How the workshop looked set to go ...
???
Lorentz workshop on music similarity, Leiden, 23 January 2015
?
My response
• How the workshop looked set to go ...
???
?
We just love questions!
Lorentz workshop on music similarity, Leiden, 23 January 2015
Categorisation without minds 1
Elements in the periodic table
Lorentz workshop on music similarity, Leiden, 23 January 2015
Categorisation without minds 2
Species
By evelynbelgium (originally posted to Flickr as IMG_6052) [CC BY-SA 2.0
(http://creativecommons.org/licenses/by-sa/2.0)], via Wikimedia Commons
Lorentz workshop on music similarity, Leiden, 23 January 2015
By evelynbelgium (originally posted to Flickr as IMG_6052) [CC BY-SA 2.0
(http://creativecommons.org/licenses/by-sa/2.0)], via Wikimedia Commons
Categorisation and similarity
• Categories correlate with similarity
– In the periodic table, similarity in electron-shell structure
– In species, in DNA sequence
• BUT simple similarity does not reliably predict
categorisation
– Different individuals of the same species have different
DNA
– Crucial factor is that recombinations of DNA lead to a
viable and fertile organism
Lorentz workshop on music similarity, Leiden, 23 January 2015
Principle
Similarity depends on the outcomes of some process.
Lorentz workshop on music similarity, Leiden, 23 January 2015
Definition of music similarity
Two instances of music are similar when there is a
plausible musical process which produces identical or
similar outcomes for the two instances.
Lorentz workshop on music similarity, Leiden, 23 January 2015
‘... a plausible musical process ...’
• Plausible:
– Tapping whenever a note begins (equivalent to transcribing
the rhythm)
• Not plausible:
– Tapping on every third note
• Musical:
– Finding out who wrote a piece of music
• Not musical:
– Selecting pieces whose composers’ names begin with ‘B’
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 2
• Almost opposite talks from Larry and Ich about
categorisation
– Larry: different kinds of similarity, some involving complex
functions
– Ich: all down to k-nearest neighbours
– Larry: unknown categories, few examples
– Ich: known categories, many examples
• Talks from Emilia and Vincent both emphasising
functions on the data
– Emilia: task dependency, examples of applications
– Vincent: DSP functions designed to stably remove
variability
Lorentz workshop on music similarity, Leiden, 23 January 2015
What in music does not involve similarity?
• Possible answers
–
–
–
–
Tuning
Instrumental technique
Notation and transcription
Some kinds of analysis (simple harmonic analysis, some
kinds formal analysis, etc.) involving recognition of preexisting categories.
– Harmonisation
Lorentz workshop on music similarity, Leiden, 23 January 2015
Conjecture
Musical processes which are not dependent on the
passage of time do not prominently involve similarity.
Lorentz workshop on music similarity, Leiden, 23 January 2015
Similarity-related Musical Processes
• Variation
– The theme, or something from the theme, is used as a
basis for composing new pieces of music
• Jazz standards
– A well known ‘song’ is performed with alterations,
improvisations, etc.
• Cover versions
– One band performs a song originated by another, but do it
in such a way as to give it a new identity
• Oral traditions
– A singer or player plays ‘the same’ song or dance but
changes are introduced, accidentally or deliberately
Lorentz workshop on music similarity, Leiden, 23 January 2015
Study the Process
• To understand similarity, perhaps we should study
the processes which give rise to it
–
–
–
–
–
Variation
Performance of jazz standards
Cover versions
Oral traditions
...
Lorentz workshop on music similarity, Leiden, 23 January 2015
Other Musical Processes
• Can still involve similarity
• Film/TV music
– Music to accompany similar segments of films
– Music in advertisements for similar products
• Playlists
– Being in the same position in a playlist is a better indicator
of similarity than just being in the same list
– Index songs by which songs they follow/precede
Lorentz workshop on music similarity, Leiden, 23 January 2015
Proposal
Find similarity data from some less obvious, but
realistic musical sources.
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 2 Working Groups
•
•
•
•
Users, individual differences and personalisation
Holistic approach, context
Scaling up
Evaluation
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 3
• Arthur: evaluation
– Need to control variables
– Are there radically different ways of doing evaluation?
• Daniel: psychological approach
– Court decisions on plagiarism as a metric
• Geraint: more psychology
– Two-way sensing-anticipating mechanism
– Continuity of representations
• Lisa: even more psychology
– Repetition is musical
– Repetition is musical
Lorentz workshop on music similarity, Leiden, 23 January 2015
Example: Schenkerian Reduction
Lorentz workshop on music similarity, Leiden, 23 January 2015
Automatically Derived Reduction
See A. Marsden, ‘Schenkerian Analysis by Computer: A Proof of Concept’, JNMR 39/3 (2010) 269-289
Lorentz workshop on music similarity, Leiden, 23 January 2015
Similarity Based on Reduction
See Keiji Hirata, Satoshi Tojo, and Masatoshi Hamanaka, ‘Cognitive Similarity Grounded by Tree
Distance from the Analysis of K.265/300’, CMMR, 2013, 415-430
Also see David Rizo, Symbolic Music Comparison with Tree Data Structure. PhD thesis, University of
Alacante, 2010
Lorentz workshop on music similarity, Leiden, 23 January 2015
Studying Reduction via Variations
See A. Marsden, ‘Recognition of variations using automatic Schenkerian reduction’, ISMIR 2010, 501-506
Lorentz workshop on music similarity, Leiden, 23 January 2015
Observation
Sometimes similarity it is easier to obtain data on
similarity than data on the musical process.
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 3 Working Groups
•
•
•
•
•
Ecological validity
Use cases
Integration of domains/dimensions/factors
Open data
Interdisciplinary collaboration
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 4
• Rolf Inge: motion cognition
– Unconscious tendency to mimic leads to motion-related
similarity
• Meinard: feature engineering
– Design the feature extraction well to make the matching easier
(cf. Vincent’s DSP approach, especially re variance)
• Johanna: performance expression
– Controlled data from multiple sources
• Gérard: creative applications of MIR
– Combination of discrete state data and continuous decaying
activation
– Evaluation of components in the lab but of the system ‘in the
field’
Lorentz workshop on music similarity, Leiden, 23 January 2015
Listening
Is this the most ubiquitous piece of musical equipment
in the world?
So what does this mean for musical tasks?
Lorentz workshop on music similarity, Leiden, 23 January 2015
Challenge
What is the task in listening to music?
Lorentz workshop on music similarity, Leiden, 23 January 2015
Day 4 Working Groups
• Relations
learning system ↔
model ↔
human (expert) knowledge/behaviour
• Going beyond few and simple features
• Computational complexity
• Brittleness on scaling up or change of context/task
Lorentz workshop on music similarity, Leiden, 23 January 2015
Tasks, tasks, tasks
By Charlotte Henard (posted to Flickr #todolist) [CC BY-SA 2.0
(http://creativecommons.org/licenses/by-sa/2.0)]
Lorentz workshop on music similarity, Leiden, 23 January 2015
Getting back to generalities?
• Humans are extremely adaptable
– But they tend to start from what they know
– So new tasks approached using the techniques of old tasks
• Tends towards overlap between behaviours/
processes
• How much overlap is there between different
musical tasks?
– Could/should we find out?
Lorentz workshop on music similarity, Leiden, 23 January 2015
Reconsideration
Perhaps general musical similarity does not really exist,
but there is enough in overlap between tasks for it to
be a useful fiction.
Lorentz workshop on music similarity, Leiden, 23 January 2015
Summary
•
•
•
•
Remember there is a world beyond ISMIR!
Music similarity is related to musical processes/tasks.
We can get similarity data from many tasks/sources.
Perhaps we should study the tasks instead of the
similarity ...
• but sometimes getting data on similarity is easier
than getting data on the task.
• The most common musical activity, listening, has no
obvious task.
• If ‘music similarity’ tout court is useful, it is because
there are commonalities between different tasks.
Lorentz workshop on music similarity, Leiden, 23 January 2015
Thank you
Beethoven, An die ferne Geliebte. Baritone,
Dietrich Fischer-Dieskau; Pianist, Hartmut Holl
Lorentz workshop on music similarity, Leiden, 23 January 2015