Information Technology Reference
In-Depth Information
Table 18.12 Assessment on the Test set. In bold, the best conflict detector
Conflict detector
# of features
UAR (%) on Test set
SO-conflict detector
315
83.4
AO-conflict detector
335
85.3
IS-2013 baseline system
6,373
80.8
Grèzes et al. ( 2013 )
1
83.1
Räsänen and Pohjalainen ( 2013 )
349
83.9
and 4.5 % (AO-conflict detector) on the Test set compared to the baseline results
with the IS-2013 set (UAR of 80.8 %) for the conflict detection task. These results
confirm also that the two types of overlap detectors ( f N, L, H g and f N, O g )are
relevant for the detection of conflict. The other results are those obtained by the
other participants. In Grèzes et al. ( 2013 ), a UAR of 83.1 % was obtained on the
Test set using a unique feature: the percentage of overlap predicted by an SVM-
based regression model. In Räsänen and Pohjalainen ( 2013 ), a UAR of 83.9 % was
obtained on the Test set using 349 relevant features selected from the IS-2013 feature
set. Feature relevance was computed by a random process. We notice that the two
better results were obtained by a similar number of features (335 vs. 349).
18.6
Conclusions
This article presents and assesses a detection system of conflict in group discussions
from voice analysis. The system was based on a multi-expert architecture and
detected two states (conflict/nonconflict). The analysis of the Train set of the SSPNet
database has demonstrated that the conflict level was highly correlated with the
mean number of interruptions, the mean duration of overlap, and the percentage
of overlap duration. The multi-expert architecture enabled knowledge regarding
overlaps to be used in the conflict detector.
The concept of LLC-Ovs and HLC-Ovs has been introduced and investigated.
Two types of overlap detectors have been developed: the first type aims at detecting
whether a speech segment contains overlap, and the second type aims at detecting
whether a speech segment contains an LLC-Ov or HLC-Ov. The accuracy of
the detectors shows that the LLC-Ovs and HLC-Ovs can be modeled. The high-
frequency mel bands and the normalized loudness are shown to be the audio
characteristics that are relevant to discriminating these two types of overlap. A
multi-resolution framework has been developed for the overlap detectors, to improve
the robustness of the detection. Three segment durations have been chosen (1, 2, and
5 s). The experiments have shown that these detectors were not redundant.
A composite set of 335 features, which consist of audio-based features and
overlap detector-based features, has been defined for the conflict detection task of
the Interspeech 2013 Conflict Challenge. The performance obtained for the Test set
 
Search WWH ::




Custom Search