Digital Signal Processing Reference
In-Depth Information
10.4.4.2 SLC Database
99 participants took part in six partial sleep deprivation studies for the recording of the
Sleepy Language Corpus (SLC) [
75
,
212
]. The mean age of subjects was 24.9 years,
with a standard deviation of 4.2 years and a range of 20-52 years. The recordings took
place in a realistic car environment or in lecture-rooms. Audio was recorded with a
sampling rate of 44.1 kHz, then down-sampled to 16 kHz; quantisation is 16 bit, the
microphone-to-mouth distance was 0.3 m.
The speech data consists of different tasks as follows: isolated vowels (sustained
vowel phonation, sustained loud vowel phonation, and sustained smiling vowel
phonation), read speech from “Die Sonne und der Nordwind” (the story of 'the
North Wind and the Sun' in German, widely used within phonetics, speech pathol-
ogy, and alike), commands and requests (10 simulated driver assistance system com-
mands/requests in German, e.g.,“Ich suche die Friesenstrasse” ('I am looking for
the Friesen street'), and four simulated pilot-air traffic controller communication
statements), and a description of a picture and a regular lecture.
A well established, standardised subjective sleepiness questionnaire measure, the
Karolinska Sleepiness Scale, was used by the subjects (self-assessment) and addi-
tionally by the two experimental assistants (observer assessment, given by assessors
who had been formally trained to apply a standardised set of judging criteria). In the
version used, scores range from 1-10: extremely alert (1), very alert (2), alert (3),
rather alert (4), neither alert nor sleepy (5), some signs of sleepiness (6), sleepy but
no effort to stay awake (7), sleepy and some effort to stay awake (8), very sleepy
with great effort to stay awake / struggling against sleep (9), extremely sleepy, cannot
stay awake (10). Given these verbal descriptions, scores greater than 7.5 appear to
be most relevant from a practical perspective as they describe a state in which the
subject feels unable to stay awake.
For training and classification purposes, the recordings (mean KSS
=
5.9, standard
deviation KSS
2.2) were divided into two classes: not sleepy ('NSL') and sleepy
('SL') samples with a threshold of 7.5 (approx. 94 samples per subject; in total 9 277
samples). A more detailed description of the data can be found in [
197
,
212
,
213
].
The available turns were divided into males (m) and females (f) per study. Then, the
turns from male and from female subjects were split speaker-disjunctive, in ascending
order of subject ID, into training (roughly 40 %), development (30 %), and test (30 %)
instances. This subdivision not only ensures speaker-independent partitions, but also
provides for stratification by gender and study setup (environment and degree of sleep
deprivation). Out of the 99 subjects, 36 (20 f, 16 m) were assigned to the training, 30
(17 f, 13 m) to the development, and 33 (19 f, 14 m) to the test set. All turns which
include linguistic cues for the sleepiness level (e.g.,“Ich bin sehr müde”—“I'm very
tired”) were removed from the test set—188 in total. The distribution of instances is
given in Table
10.26
.
=
