Digital Signal Processing Reference
In-Depth Information
Table 11.14
WA per feature type for data-driven SVMs (no gain was reached by optimisation of
the feature space)
WA [%]
Key
Sub
Dom
Sum
Min
Maj
Sum
All
62.1
4.0
6.7
72.8
2.9
9.2
84.9
CHROMA
59.6
4.6
7.9
72.1
1.2
11.0
84.3
Scale
43.3
10.4
13.1
66.8
0.0
10.0
76.8
Scale dom
42.1
11.3
12.7
66.1
0.0
9.6
75.7
Scale cad
40.0
11.9
14.0
65.9
0.0
8.8
74.7
Chords
49.4
6.5
11.2
67.1
0.0
13.3
80.4
Chords dom
46.2
8.7
12.1
67.0
0.0
11.0
78.0
Chords cad
41.9
11.2
13.5
66.6
0.0
9.6
76.2
PTR maj 48.8 6.3 11.9 67.0 0.0 13.5 80.5
PTR maj dom 46.3 8.3 11.9 66.5 0.0 10.8 77.3
PTR maj cad 42.5 9.2 12.5 64.2 0.0 9.4 73.6
PTR min 54.8 4.4 8.5 67.7 0.0 14.6 82.3
PTR min dom 53.3 4.4 10.0 67.7 0.0 14.2 81.9
PTR min cad 45.4 8.7 12.7 66.8 0.0 10.2 77.0
correct key (Key) and percentage of confusion with (sub-)dominant (Sub/Dom), and relative minor
and major (Min/Maj). Database KEY-ALL, ten-fold SCV, 24 keys
Table 11.15
WA for correlation ('PTR maj/min dom' features) versus data-driven SVMs ('all'
features, ten-fold SCV), subdivided per genre
WA [%]
Key
Sub
Dom
Sum
Min
Maj
Sum
Correlation
MTV
46.5
5.0
11.5
64.0
14.0
0.5
77.5
CHANSON
66.4
8.0
10.8
85.2
0.0
2.0
87.2
CLASSIC
76.4
0.0
4.5
80.9
5.6
1.1
87.6
JAZZ
34.1
6.1
13.4
53.7
17.1
2.4
73.2
KEY-ALL
55.4
5.2
10.4
71.0
9.0
1.3
81.3
Data-driven SVMs
MTV 52.0 3.0 9.5 64.5 7.0 9.0 80.5
CHANSON 67.8 8.7 12.8 89.3 0.0 0.0 89.3
CLASSIC 68.5 10.1 4.5 83.1 3.4 6.7 93.2
JAZZ 46.3 8.5 7.3 62.1 9.8 6.1 78.0
KEY-ALL 62.1 4.0 6.7 72.8 2.9 9.2 84.9
WA for the correct key (Key) and percentage of confusion with (sub-)dominant (Sub/Dom), and
relative minor and major (Min/Maj), 24 keys
the correct key in all cases but the CLASSIC set. The JAZZ set benefits most from this
trend. Interestingly, WA is more balanced across genres for the data-learnt method.
11.4.6 Summary
Within this section performance of musical key determination on originally recorded
and MP3 encoded popular, Chanson, Classical, and Jazz music was demonstrated.
