Transcript 1 Automatic Music Style Recognition

Automatic Music Style
Recognition
Arturo Camacho
1
Contents






2
Goal
Definitions
Method
Results
Conclusions
Further work
Goal

3
Classify songs given as MIDI files into the
Latin styles Salsa, Merengue and Cumbia.
Definitions


4
MIDI file: file that represents
a song as a sequence of
musical notes and the time
at which they must be
played.
Audio file: file that
represents a sound as a
sequence of numbers that
correspond to samples of
the corresponding wave.
Method

Method is based on the following
observations:
–
–
5
Latin Rhythms have very well defined patterns for
most instruments.
These patterns are highly repetitive. They are
present most of the time during a song.
Method



6
Make templates of the rhythms.
Compare songs with templates.
Measure the degree of similarity and decide
about the membership of the song to the
style.
Making Templates



7
Listen to many songs with the same style
and identify patterns for each of the
instruments.
Identify similar patterns and group them
together into a template.
Make variations of a pattern optional
elements.
Example of Template

Salsa common bass
patterns:
“3-2 Clave”
“2-3 Clave”

Template for bass rhythm:
( )
–
Compiled into one measure:
( )
8
Comparing song with Template

Compare your song with the
template and sum one point
for every match.
–

9
Example:

Your song:

Template:
If the score is higher than
some predetermined
threshold accept the song,
otherwise reject it.
( )
+1+1 +1
( )
+1 +1+1 +1+1 +1+1
+1 +1+1
if 13 > threshold  accept
if 16
Using more than one instrument



10
Make a template for each of the instruments.
Compare each of the instruments in your
song with the corresponding template.
Accept the song if the number of accepted
instruments is higher than some
predetermined threshold.
Using more than one style



11
Make templates for each of the styles.
Compare your song with each of them and
list all the styles that accepted the song.
If more than one style accepted the song,
take the style with highest score.
Results



12
Method was implemented to identify Salsa,
Merengue, and Cumbia based on bass and
congas.
It was tested with 9, 12, and 12 songs of
each style, respectively.
88% of the songs were correctly classified.
Conclusions


13
The method described is based on the fact
that most of music is highly repetitive. Given
this, templates for each style were identified
and methods to compare a song with such
templates were developed.
It showed to be very effective. In the
experiments it classified correctly 88% of the
songs.
Further work:
Latin Rhythms Identification on AUDIO files

Initial approach: identification based on bass line
–
–
Lowest frequency component is usually played by the bass.
Method:




Applications:
–
14
Extract the lowest frequency component and try to identify the bass
notes.
Recognize beats.
Apply same methodology as before.
Search of music by style in multimedia databases, personal
collections, and audio devices.