Automatic Analysis of Music Content

1
Automatic Analysis of Music Content

• tepán Albrecht
• Department of Computer Science
• Faculty of Applied Sciences
• University of West Bohemia in Pilsen
• 31.3.2008

2
Motivation

• Thousands hours of music in the internet and
  various archives
• Need to
• retrieve information from the music content
  Music Information Retrieval (MIR)
• To retrieve we need music content description
• Quest to do it automatically
• MIR - significant promise for the future
  development of commercial and research
  applications such as
• searchable databases
• popular tunes on the web and databases of
  historical music
• score typesetting programs

3
Problem Definition

• Automatic analysis of music content
• Researchers and developers focus on retrieval of
• notes in a polyphony (chords)
• notes of a melody, bass and drum line in a given
  song
• separable perceptual sound sources
• genre of a given song (rock/pop/classic/)
• structure of a song (verse, bridge, chorus,..)
• tempo and meter key of a song
• music instrument name in a sound excerpt
• similarity of two songs

4
Problem Definition - Limits

• Music ambiguity
• 2 complex sounds can appear as a different single
  complex sound
• Still a skilled musician is better in music
  content recognition than the best music content
  analysis software

5
Approaches to The Solutions
6
The Project - Motivation

• Directly, tries to solve together
• tracking of notes in a polyphony music
• sound source separation problem
• recognition of the inner sounds ( instrument
  identification)
• Common approaches
• train models for recognition of a few musical
  instruments
• Modern music
• not composed by a few clean intrument sounds, but
  also by sounds recorded in miscellaneous
  ambiences or by artificial sounds
• This concept is designed so as to work without a
  limit on sounds in general

7
The Project - Definition

• Called Music Signal Decomposition
• Mono audio signal (a song)
• composition of anything - from the tone A of a
  piano to the drum loop in a popular song
• Set of audio components
• Sound represented by a component
• Tries to identify their presence and form in the
  song

8
The Project Methods To Solve

• Complex Task
• Bayesian Statistical Methods (BSM)
• their framework allows to utilize any helpful
  information (heuristic),
• e.g., about number of polyphony, about position
  of component, about its form,
• Information represented by a probability density
• Powerful framework for the solution of complex
  problems
• besides audio signal analysis, they have proven
  their efficiency in image analysis, target
  tracking (theory control), genetics and financial
  analysis
• Exact formula estimating the wanted parameters
  leads to intractable inthegrals, we utilize its
  approximation - Monte Carlo methods

9
My Experience with MIR

• Devote myself to it since Erasmus stay in
  Joensuu, Finland (2004)
• 2005 MSc. Thesis
• retrieval of notes in polyphonic music of
  harmonic instruments
• Since autumn 2005 the Ph.D. student
• Articles dealing with the Audio Signal
  Decomposition
• Music Signal Decomposition Based on
  Identification and Subtraction of Components -
  DAGA 2008, German Acoustic Society Conference,
  Dresden, March 2008
• Music Signal Decomposition Approaches Based on
  Unsupervised Source Separation Methods - PhD
  Workshop, Hungary, Balatonfuhred, September 2007

10
Why to join prof. Phillipe Preux research group

• In autumn 2007 I was as the intern at prof. Preux
  research group. That stay helped me essentially
  with the same project understanding and designing
  the first models of solution
• The people from that research group deal with
  Bayesian Statistical Methods
• No research group for Music Information Retrieval
  in the Czech Republic
• I apply for
• 3 months scholarship
• 6 months of doctorat en cotutelle

11
The End

• Thank you for your attention.