Vision - Intelligent Audio Analysis - page 328

Table 14.2 continued
Task #Classes Database #Train #Test Test method UA [%] WA [%]

Music classes
Metre

2 BRD 1855 1855 SCV 97.1 96.6

Dance style 9 BRD 1855 1855 SCV 88.9 89.1

Tempo 141 BRD 1855 1855 SCV 89.0 88.5

Key 12 KEY-ALL 521 521 SCV - 77.3

Key 24 KEY-ALL 521 521 SCV - 62.1

Chords 24 ChoRD 10702 10702 LOSO - 60.13

Chords 36 ChoRD 10702 10702 LOSO - 48.84

Mood-arousal 3 NTWICM 1376 1272 T/D/T 56.2 58.7

Mood-valence 3 NTWICM 1376 1272 T/D/T 61.2 61.0

Voice presence 2 UltarStar 326527 97144 T/D/T 75.77 75.81

Singer age 2 UltraStar 315043 93342 T/D/T 57.55 56.56

Singer gender 2 UltraStar 326198 96373 T/D/T 89.61 93.60

Singer height 2 UltraStar 280714 80962 T/D/T 72.07 78.26

Singer race 2 UltraStar 321178 96563 T/D/T 63.30 76.98

Sound classes
Birds 2 HU-ASA 868 868 SCV 80.0 81.3
Animals 5 HU-ASA 1063 1063 SCV 49.5 64.0
Acoustic events 7 FindSounds 11292 5645 T/D/T 66.5 71.7
Given are the number of classes, the database with training (uniting training and development
instances) and test instances and the test method, where T/(D/)T are train, (develop), and test and
SCV is always ten-fold. The level of precision of the presented results depends on the number of
test instances

Achallenge remaining at that point will be the careful evaluation of ethical issues if
machines can listen to and understand arbitrary audio including personal information
and details.
Finally, given such holistic analysis capability basing on very efficient source sep-
aration and synergistic coupling of tasks, future audio analysis systems can start to
train themselves in a massive way such as by crawling the Internet for audio, or lis-
tening to very general media broadcast potentially reaching supra-human capabilities
in some of the alluded tasks.

Reference

1. Schiel, F.: Perception of alcoholic intoxication in speech. In Proceedings of Interspeech,
pp. 3281-3284. Florence (2011)