Digital Signal Processing Reference
In-Depth Information
•
structural (behavioural, interactional, social) signals
: role in dyads, groups, and
the like [
201
], friendship and identity [
202
], positive/negative attitude [
203
], (non-
verbal) social signals [
204
], entrainment [
205
].
Two such were picked out as a central theme in another follow-up challenge
which focused on the crucial application domain of security and safety: the com-
putational analysis of intoxication and sleepiness in speech. Apart from intelligent
and socially competent future agents and robots, main applications are found in the
medical domain and surveillance in high-risk environments such as driving, steering
or controlling [
206
].
In [
207
], several differences are shown in the quality of the vocal articulation after
a night of sleep deprivation (reduced intonation and a slowing down of the vocal flow);
in [
208
], a reduction of the spontaneous dialogues and performance degradation of
the subjects is observed under similar conditions. Generally speaking, these results
suggest effects of sleep deprivation on communication, especially with a reduction of
the spontaneous verbalisations, trouble finding words, and a degradation of the artic-
ulation. Subjects under sleep deprivation produce less details and show less empathy
toward a team-mate [
209
]. Some stressors such as alcohol are likely to influence
articulators, which helps to explain intra-speaker and inter-speaker variability [
210
].
For the experimental evaluation of these tasks, the Alcohol Language Corpus
(ALC) and the Sleepy Language Corpus (SLC) with genuine intoxicated and sleepy
speech were provided [
77
]. The first consists of 39 h of speech from 154 speakers
in gender balance. It serves to evaluate features and algorithms for the estimation
of speaker intoxication in gradual blood alcohol concentration (BAC). The second
features 21 h of speech recordings of 99 subjects, annotated in the 10 different levels
of sleepiness of the Karolinska Sleepiness Scale (KSS) [
211
].
The verbal material is of different complexity reaching from sustained vowel
phonation to natural communication. In part, the corpora feature detailed speaker
meta data, orthographic transcript, phonemic transcript, segmentation, and multiple
annotation tracks. Again, both were given with distinct definitions of test, develop-
ment, and training partitions, with a strict speaker independence as needed in many
real-life settings. Two tasks are addressed:
First, the alcoholisation of a speaker is determined as two-class classification task:
alcoholised
for a BAC exceeding 0.5 per mill
12
or
non-alcoholised
for a BAC equal
or below 0.5 per mill. The measure of interest is—as before—UA of these two classes
to better compensate for imbalance between classes.
Second, the sleepiness of a speaker is determined by a suited algorithm and
acoustic features. While the annotation provides sleepiness from 1-10 on the KSS,
only two classes are recognised: sleepiness for a level exceeding 7.5 on the KSS, and
non-sleepiness for a level equal or below 7.5. Again, the measure is UA of the two
classes and a further enlarged standard feature set is used [
77
].
12
Per mill BAC by volume (standard in most central and eastern European countries; further ways
exist, e.g., percent BAC by volume, i.e., the range resembles 0.028 to 0.175 per cent (Australia,
Canada, USA), points by volume (GB), per mill by BAC per mass (Scandinavia) or part per million.)
