Issues in Automatic Musical Genre Classification

1
Issues in Automatic Musical Genre Classification

• Cory McKay

2
Introduction to musical genre

• Practical importance
• Radio stations
• Libraries
• Retailers
• Theoretical importance
• How we construct genre taxonomies
• Mechanisms we use to classify music
• Mechanisms we use to distinguish between
  categories

3
Introduction to musical genre

• No universally accepted set of categories
• Genre descriptions are rarely consistent,
  comprehensive, clear or objective
• Genre constructed by a complex interaction of
• Marketing strategies
• Historical conventions
• Choices made by music librarians, critics and
  retailers
• Interactions of groups of musicians and composers

4
Introduction to musical genre

• Difficulties with classifying by musical genre
• What categories should be used?
• What are the boundaries between categories?
• How are different categories related to each
  other?
• What are characteristics of a particular genre?
• What genre(s) do individual pieces belong to?
• Genres constantly changing and being created
• Main problems in automatic genre classification
• Which features should be used?
• What taxonomy should be used?

5
Symbolic vs. audio representation

• Using a symbolic representation of music rather
  than an audio representation
• Allows one to think of music in terms of musical
  features rather than signal processing features
• Also allows one to classify scores for which no
  audio recordings are available
• Future advances in automatic transcription
  systems will allow use of both types of features

6
Feature extraction

• Features
• Characteristic pieces of information that can be
  extracted from music and used to describe or
  classify it.
• Features are very important, as they are the only
  percepts available to classification systems.
• Want features that demonstrate differences
  between categories.

7
Feature extraction

• Sophisticated theoretical analyses
• Too genre-specific
• Automatic analysis often an unsolved problem
• Want features that can be represented as simple
  numbers that can be fed to classification system.
• Want features with musicological meaning if
  possible.
• Want a large catalogue of features so that
  classifier can choose ones best suited to
  particular types of classification and
  sub-classification (hierarchal).

8
Feature extraction

• Have devised 160 features based on
• Instrumentation
• Texture
• Rhythm
• Dynamics
• Pitch Statistics
• Melody
• Chords
• Scope of these features not limited to genre
  classification.
• Could be used for a variety of classification,
  clustering and analysis tasks.

9
Feature extraction

• Future research extract non-musical features
• Lyrics
• Clothing
• Album art
• etc.
• Research of Whitman Smaragdis (2002) a good
  start in this direction.

10
Automatic classification techniques

• Three main automatic classification paradigms
• Expert Systems Use pre-defined rules to process
  features and arrive at classifications.
• Require explicit a priori knowledge of rules
• Great deal of effort required to change once
  implemented
• Unsupervised Learning Cluster the data based on
  similarities that the systems perceive
  themselves. No model categories are used.
• Categories generated objective and not likely
  to correspond to categories used by humans

11
Automatic classification techniques

• Supervised Learning Attempt to formulate
  classification rules by using machine learning
  techniques to train on model examples. Previously
  unseen examples are classified into one of the
  model categories using the patterns learned
  during training.
• Require a pre-defined taxonomy and pre-classified
  training examples
• Supervised learning is the best option for the
  particular problem of genre classification
• Several possible implementations nearest
  neighbor, neural networks, induction trees, etc.

12
Forming genre taxonomies

• Using hierarchal taxonomy allows the inclusion of
  both broad and specialist categories
• Could devise rational but artificial categories
• Not realistic and therefore not useful
• Experimental approach
• Music industry categories (e.g. Billboard,
  Grammies, etc.)
• Specialty shows on TV and radio
• Specialist interviews (DJs, music reporters,
  etc.)
• Retailers, including on the Internet

13
Forming genre taxonomies

• Data-mining techniques
• Computers automatically search text resources on
  the web and attempt to form categories and
  correlations
• Holds a great deal of potential, but is difficult
  to implement and still untested

14
Existing automatic genre classification systems

• Most experiments to date have been with audio
  rather than symbolic recordings
• Success rates of between 61 to 93 when dealing
  with between 3 and 10 categories
• Only a few studies of symbolic classification
• 63 to 84 for between 2 and 3 categories

15
Classification experiment

• Did initial experiment to test viability of
  symbolic classification
• Used 225 MIDI recordings divided into 3 parent
  genres and 9 sub-genres
• Classical
• Baroque, Romantic, Modern
• Jazz
• Swing, Cool, Funky
• Pop
• Rap, Country, Punk
• Categories were just roughly chosen for test
  purposes
• Could just have easily used other formats, such
  as Humdrum or GUIDO

16
Classification experiment

• Performed classification with 8 neural networks
  and a coordination system.
• Factors increasing difficulty
• Only 20 recordings per genre were used for
  training for each run to represent a wide range
  of musics within each category increased
  difficulty.
• Only 20 features were implemented

17
Classification experiment

• Average success rates
• 85 for parent genres
• 58 for sub-genres
• Results were fairly consistent across training
  runs
• These rates comparable to existing audio
  classification systems using similar numbers of
  categories and better than existing systems using
  symbolic data.
• Encouraging

18
Classification experiment
19
Classification experiment

• Future improvements
• Use realistic taxonomy
• Larger training and sample set
• More features
• More sophisticated classification methodology
• Feature selection sub-system for each level of
  classification hierarchy

20
Software interface

• A user-friendly interfaced is being developed
  that will be ported to the classification system.
• Easy to use and flexible so that it can be used
  for a variety of research and applied purposes by
  people with little technical expertise.
• Allows user to
• Input and edit arbitrary taxonomies and lists of
  recordings
• Choose which features to extract
• Evaluate the usefulness of particular features in
  different contexts
• Evaluate effectiveness of different
  classification techniques
• Additional features can be designed and added to
  the software easily and painlessly by anyone with
  some basic Java programming skills.

21
(No Transcript)
22
(No Transcript)
23
Conclusions

• Could use system such as this to study
• Particular taxonomies
• How well different features perform in different
  contexts
• Differences between and definitions of particular
  genres
• Could easily adapt system to other types of
  classification
• Composer / performer
• Historical / geographical / cultural
  characteristics
• Personal preferences
• Practical applications
• On-line musical databases of any kind

24
Questions