Digital Signal Processing Reference
In-Depth Information
Table 11.12
WA per feature type for data-driven SVMs: correct key (Key) and percentage of
confusion with (sub-)dominant (Sub/Dom)
WA [%]
Key
Sub
Dom
Sum
All groups
76.2
7.3
9.8
93.3
Optimised space
77.3
7.1
9.8
94.2
CHROMA
73.7
7.5
11.2
92.4
Scale
59.6
13.3
18.1
91.0
Scale dom
55.2
15.0
19.6
89.8
Scale + cad
51.5
17.1
20.8
89.4
Chords
68.8
9.6
12.9
91.3
Chords dom
65.0
10.2
15.6
90.8
Chords + cad
54.4
15.6
20.4
90.4
PTR maj
68.5
8.8
13.8
91.0
PTR maj dom
63.7
9.0
16.5
89.2
PTR maj cad
56.7
14.4
19.0
90.1
PTR min
73.5
7.7
11.3
92.5
PTR min dom
72.5
8.5
11.5
92.5
PTR min cad
62.5
11.3
17.1
90.9
Database KEY-ALL, ten-fold SCV, 12 keys
Table 11.13
WA for correlation ('scale cadence' features) versus data-driven SVMs ('optimised
space', ten-fold SCV) per genre
WA [%]
Key
Sub
Dom
Sum
Correlation
MTV
74.5
9.0
8.5
92.0
CHANSON
70.5
12.8
9.5
92.8
CLASSIC
86.5
6.7
3.4
96.6
JAZZ
68.3
1.2
18.3
78.8
KEY-ALL
72.3
7.5
12.7
92.5
Data-driven SVMs
MTV
73.0
8.0
10.0
91.0
CHANSON
72.5
8.1
10.1
90.7
CLASSIC
82.0
6.7
5.6
94.3
JAZZ
59.8
14.6
17.1
91.5
KEY-ALL
77.3
7.1
9.8
94.5
Correct key (Key) and percentage of confusion with (sub-)dominant (Sub/Dom), 12 keys
WA drops by roughly 15 % absolute to 62.1 % at maximum.Considering that pieces
in major keys make up 71.5 % of the data, leaving only 28.5 % for minor keys, may
explain the majority of confusions being in favour of relative major keys and almost
none the other way round. This could be overcome by balancing. In this respect,
interestingly, balancing by cyclic key-shift did not improve results [
28
].
In Table
11.15
a comparison is made for 24 keys as previously between the two
assignment approaches in optimal configuration, each: 'PTR maj/min dominant'
features for correlation and 'all' features for SVMs. SVMs prevail over correlation for
