Applications in Intelligent Music Analysis - Intelligent Audio Analysis - page 250

Table 11.10 WA for different 'gating' lengths from the beginning of a musical piece

WA [%] First 30 s First 60 s First 90 s First 120 s Whole piece

MTV 59.5 70.0 71.5 69.0 70.5

CHANSON 72.5 75.8 75.2 71.1 72.5

CLASSIC 62.9 76.4 75.3 74.4 82.0

JAZZ 37.8 50.0 51.2 48.8 59.8

KEY-ALL 66.2 74.2 75.8 75.2 75.6

SVM, ten-fold SCV, range C3-C8, 12 keys

Table 11.11 WA of different frequency ranges. SVM, ten-fold SCV, 12 keys

WA [%] C2-C6 C2-C7 C2-C8 C2-C9 C3-C7 C3-C8 C3-C9

MTV 58.0 63.5 67.5 67.0 64.0 70.5 69.5

CHANSON 57.0 61.7 71.8 71.8 59.7 72.5 71.1

CLASSIC 70.8 79.8 80.9 82.0 73.0 82.0 80.9

JAZZ 45.1 48.8 61.0 57.3 51.2 59.8 56.1

KEY-ALL 66.0 69.4 75.6 75.0 70.7 76.2 75.0

the corresponding correlation template vector ( 11.30 ). An example in the key of C
major (as before) is shown—symbols are explained in Eq. 11.32 ) where maj / min /
cad / dom abbreviate major / minor / dominant / cadence as before. For the key K
holds:

T

K = arg max

k

κ · t cor ( k ) (11.28)

x

p min , cad
(11.29)

κ ∈ , s , s dom , s cad , c , c dom , c cad , p maj , p maj , dom , p maj , cad , p min , p min , dom ,

) = 0 0 T

t cor ( C , 0 , 0 , 1 , 0 , 0 , 0 , 0 , 0 , 0 , 0 , (11.30)

The elements of t cor ( weight the semitones of the feature vector, and the indices i

correspond to the scale with (i = 1

k )

= = =

G#), and k corresponds
to the root for which the correlation result is calculated. The root element's value
is set to 1, and all remaining ones to zero as seen in Eq. ( 11.30 ). The key with the
highest correlation value is then decided for. Thus, the maximum component of the
input feature vector is searched for. Results per proposed feature type are given in
Sect. 11.4.5 .
A visualisation of the principle of the derived dominant and cadence features as
were introduced in Sect. 6.2.2 is found in Fig. 11.14 exemplified by Charles Trénet—
“ La mer ” in the key of C major of the CHANSON dataset: Shown is the distribution of
correlation results for the basic feature 'scale' , the derived features 'scale dominant' ,
and 'scale cadence' in the circle of fifths. As can be seen, with increasing relation
to the original key, the correlation results monotonously increase. Figure 11.14 a

A,i=2 A#,..., i = 12