Robust Polyphonic Music Retrieval with N-grams

1
Robust Polyphonic Music Retrieval with N-grams

• SHYAMALA DORAISAMY
• Department of Computing, South Kensington
  Campus, Imperial College London, London SW7 2AZ,
  UK

• STEFAN RÜGER
• Department of Computing, South Kensington
  Campus, Imperial College London, London SW7 2AZ,
  UK

2
Outline

• Introduction
• A technique for indexing polyphonic music data
• Error models
• Experimental setup
• Results

3
Introduction

• We restrict our focus to music which sound event
  information is encoded.
• Most current music IR system?
  query-by-example? query-by-humming
• Our aim is the development of a fault-tolerant
  polyphonic Music IR system.
• A known problem pertaining to IR system is the
  lack of query precision.

4
A technique for indexing polyphonic music data

• Patter extraction
• N-gram the sequence of symbols is divided
  into overlapping constant-length
  subsequence.
• We encoded polyphonic music pieces as an ordered
  pair of onset time and pitch.

5
(No Transcript)
6

• Pitch
• Intervali Pitchi1 Pitchi
• ? -2 -1, -1 1, 1 -121 3,
• Rhythm
• I1R1In-2Rn-2In-1
• ?-2 1 -1, -1 1 1, 1 1 -121 1 3

7

• Restriction to envelopes
• We suggest restricting the possible combinations
  to variations of upper and lower envelopes of the
  window.
• Pattern encoding
• C(I) int (X tanh (I / Y))
• X has the effect of limiting the number of codes
• Y determines the rate at which class sizes
  increase as interval sizes increase.

8
(No Transcript)
9
(No Transcript)
10
Error models

• Query by humming
• Expansion
• Compression
• Repetition
• Omission
• Query by example
• NewIntervalk Intervalk Di e
• NewRatiok Ratiok exp( Dr e)

11

• N-grams and fault-tolerance
• Top 0 7 0, 7 0 2, 0 2 0, 2 0 -2, 0
  -2 0, -2 0 -2, 0 -2 0, -2 0 -1, 0 -1 0,
  -1 0 -2, 0 -2 0, -2 0 2, 0 2 -4
• Bottom 0 6 0, 6 0 3, 0 3 0, 3 0 -2,
  0 -2 0, -2 0 -2, 0 -2 0, -2 0 -1, 0 -1
  0, -1 0 -2, 0 -2 0, -2 0 -2

12
Experimental setup

• Test collection
• Use 6366 polyphonic music pieces in the MIDI
  format to construct different experimental index
  mechanism.
• Known-item search and simulated errors
• Known-item search
• Quality of the retrieval mechanism
• Monophonic queries error probability p
• Polyphonic queries Eqs.(5) and (6).

13

• Ad-hoc user queries and relevance judgments
• Hand-crafted ten monophonic queries
• Subject each query ten times independently to
  McNabs error model with error probability for
  each note of 20.

14
(No Transcript)
15

• Index file development
• P3, P4 pitch-only, n3 or n4, Y24
• PR3, PR4 pitch and rhythm, n3 or n4,Y24
• CP1 Y48 for a 2-1 mapping
• CP2 Y72
• CR the encoding for the ratios, we use the codes
  A-D, Y and a-d, y
• ENV the generation of n-grams is restrict to the
  variations of the upper and lower envelopes of
  the music

16

• AL1 like PR4 but n-grams are constructed from
  every other onset time.
• AL2 like AL1, but by skipping two onsets.
• AM like PR4, but only the monophonic path
  through the highest notes are kept.

17
(No Transcript)
18
Results

• Monophonic queries and QBH
• Polyphonic queries and QBE
• Ad-hoc queries
• Precision-at-15
• W.A.

19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
Interval Win1 -2, -1 Win2 -1, 1 Win3 1
-121 3
Ratio Win1 1 Win2 1 Win3 1, 1
23

• ????????????

-5 lt z lt 5