Detecting Speech Interruptions for Automatic Conflict Detection - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 378

Table 18.5 Accuracy rates of the detectors f N, L, H g on the Development set

according to the feature sets. In bold, the best detector
Detectors Feature set N-Acc. (%) L-Acc. (%) H-Acc. (%) UAR (%)

f N, L, H g _1 IS-2010 78.0 31.7 73.5 61.1

f N, L, H g _1 IS-2011 79.9 32.7 70.5 61.0

f N, L, H g _1

IS-2012 79.4 31.4 71.4 60.7

f N, L, H g _2 IS-2010 79.5 32.6 71.2 61.2

f N, L, H g _2 IS-2011 78.1 35.9 70.0 61.3

f N, L, H g _2 IS-2012 80.5 31.5 68.0 60.0

f N, L, H g _5

IS-2010 77.5 44.7 68.3 63.5

f N, L, H g _5 IS-2011 76.4 40.0 67.7 61.4

f N, L, H g _5 IS-2012 80.8 38.2 67.4 62.1

we chose to use only the best three-class classifier: f N, L, H g _5 with the IS-2010
audio feature set. Our assumption is that only the best overlap classifier is relevant
for the detection of conflict.

18.4.4 Audio Characteristics of Overlaps

Previous studies (Smolenski and Ramachandran 2011 ; Shokouhi et al. 2013 )have
shown that the audio characteristics of overlapping speech are different from speech
in which a lonely speaker occurs. We looked for the discriminating cues (1) between
Ov and Non-Ov and (2) more specifically between HLC-Ov and LLC-Ov. For these
investigations, we chose to study the segments that had a 5-s duration in the Train
set for the best accuracy results of the 5-s-based f N, O g and f N, L, H g detectors
(see, respectively, Tables 18.4 and 18.5 in Sect. 18.4 ). The 38 low-level descriptors
(LLDs) of the IS-2010 feature set have been used as audio characteristics. The
relevance of the LLDs was analyzed with respect to the classes Non-Ov/Ov, which
are referred to as f N, O g , and the HLC-Ov/LLC-Ov, which are referred to as
f H, L g . For each LLD, the relevance is given by the information gain (Rauber and
Steiger-Garcao 1993 ), which is computed on the segments of 5-s duration with the
following formula: H (class) H (class/LLD), where H is the Shannon entropy. Four
steps were defined to compute the entropy: (1) filtering of the IS-2010 features
according to a given LLD, (2) clustering of the segments of the Train set using
the filtered features, (3) computation of the contingency table from the class and
the cluster associated with each segment, and (4) estimation of the entropy from the
table of contingency. Table 18.6 gives the information gain computed on the Train
set of the five best-ranked LLDs (over 38 LLDs) in discriminating LLC-Ovs and
HLC-Ovs. The most relevant LLDs are the logarithmic powers of mel-frequency
bands and, more precisely, the high-frequency bands and the normalized loudness.
These results show that various acoustic differences exist between the two types of
overlaps.