Applications in Intelligent Sound Analysis - Intelligent Audio Analysis - page 298

As classifier, Random Forests as ensemble of decision trees are used. This choice is
motivated by their good ability to cope with large feature spaces, as feature sub-spaces
are randomly assigned to the trees in the forest. A good configuration proved to be 30
trees, and 150 randomly assigned features for each tree. For further reproducibility
besides using an open-source feature extractor and the FindSounds database (cf.
Sect. 5.3.3 ) that can be retrieved from the Internet, the classifier implementation
provided by the Weka toolkit [ 22 ] is chosen again.

12.2.2 Performance

Considering the imbalance of instances among the classes, UA will be the mea-
sure of primary interest. Further, WA is partly provided in addition, as well as recall,
precision, and F 1 -measure. The experiments base on random partitioning of the Find-
Sounds database into three stratified folds to provide two training and one completely
disjoint testing set. The first fold (F1, 5 646 instances) is always used with its original
manually assigned labels for training. The second fold (F2, 5 646 instances) is used
either without its original labels (F2 U ) or with these labels (F2) to be able to compare
to using this fold in a semi-supervised or supervised manner for training. The third
and last fold (5 645 instances) is always used for testing. Random partitioning is
carried out with Weka's default random seed.
Table 12.5 shows the occurred confusions for seven categories of sound event
classification using the original labels training on fold one and two and testing on
the third fold. This is the 'best case' given the entirely supervised learning with
utmost data and serves as upper benchmark. Most confusions can be explained well
by common sense, such as those of sounds from people with sounds of animals or
sounds from vehicles with sounds of noise makers.

Table 12.5 'Best case' confusions when automatically classifying seven sound categories on the
FindSounds database with original labels for both training folds F1 and F2 (cf. line 'supervised
F1

+

F2' in Table 12.6 )
Truth [#]

Classified as
People Animals Nature Vehicles Noisemakers Office Instruments

People 564 153 11 26 17 25 50

Animals 126 717 7 35 23 20 18

Nature 18 35 157 42 44 10 6

Vehicles 37 37 26 476 86 15 45

Noisemakers 22 43 36 77 372 72 48

Office 29 37 1 16 111 364 31

Instruments 32 33 6 31 47 16 1 395

Confusions 264 338 87 227 328 158 198