Information Technology Reference
In-Depth Information
Fig. 18.10
Architecture scheme of the SO-conflict detector
Fig. 18.11
Architecture scheme of the AO-conflict detector
(1, 2, and 3) that correspond to the three multi-resolution-based f N, O g detectors
and a conflict detector (4). The IS-2010 feature set (1,582 features) was used
for the overlap detectors. The Cf-1 feature set (315 features) was associated with
the conflict detector. The Cf-1 feature set was obtained by a backward selection
algorithm from the Ov-1 feature set (28 features) and the IS-2103 set (6,373
features).
Figure 18.11 resumes the architecture characteristics of the second conflict
detector, called the advanced overlap-based conflict detector (AO-conflict detector).
This detector was based on a set of overlap detectors (1, 2, 3, and 4) and a
conflict detector (5). The IS-2010 audio feature set (1,582 features) was used for
the overlap detectors. The Cf-2 feature set (335 features) was associated with
the conflict detector. The Cf-2 feature set was obtained by a backward selection
algorithm from the Ov-2 feature set (55 features) and the IS-2103 set (6,373
features).
18.5.5
Conflict Detection on the Test Set
The Test set of the Interspeech 2013 Conflict Challenge (Schuller et al. 2013 )
consisted of 397 clips with no information or metadata available. Table 18.12 gives
the results obtained on the Test set during the Conflict Challenge. Experiments gave
a UAR of 83.4 % for the SO-conflict detector and a UAR of 85.3 % for the AO-
conflict detector. These results show an improvement of 2.6 % (SO-conflict detector)
Search WWH ::




Custom Search