Voice Separation A Local Optimization Approach Jurgen Kilian Holger H. Hoos

advertisement
Voice Separation
A Local Optimization Approach
Jurgen Kilian
Holger H. Hoos
Xiaodan Wu
Feb. 26 2003
Introduction


What is Voice Separation? (score)
Some of the Usages


To obtain usable scores from performances of
polyphonic music
To improve the music retrieval systems that
only support monophonic music
Existing Approaches


Split Point Separation (figure)
Rule Based Approaches




Prefer small intervals between succeeding
notes
Keep range of a voice small
Use a small number of voices
Avoid crossings of voices
Overview
The method developed by the authors


Goal: To create reasonable and flexible
score-notation for various needs
Technique and Algorithm



Preprocessing
The Cost Function
Stochastic Local Search Approach
Preprocessing


Removes small overlaps, quantizes
the notes
Partitions the musical piece into
slices.
The Cost Function

Pitch Distance Penalty
Penalize large pitch intervals between successive notes
in a voice.

Gap Distance Penalty
Penalize large gaps/rests between successive notes in a
voice.

Chord Distance Penalty
Penalize chords with a large pitch interval between the
highest and the lowest note in a voice, as well as irregular
chords.

Overlap Distance Penalty
Penalize overlaps between successive notes in the same
voice.
Cost-Optimized Slice Separation


Using a stochastic local search
approach
A fixed number of steps was employed
to terminate the improvement
  3  yi  nVoice
Implementation and Results





Implemented in midi2gmn, a program for
converting MIDI into GUIDO Music Notation.
The code was written in ANSI C++ .
All the penalty parameters and the maximum
number of voices could be set in an initialization
file.
With the correct parameter settings, the tested
Bach chorals and inventions were separated
almost entirely correctly.
Drawback: If a voice continues with a large
interval step after a rest, the algorithm acts
incorrectly.
Comments on the future work


Cost Functions could be improved
by employing more rules in music
theory.
The thought of Expectancy could be
combined with the cost functions.
score
Split point separation
Download