Digital Signal Processing Reference
In-Depth Information
900
600
600
450
F1 - Females
F1 - Males
300
300
Passenger
Conversation
American
Airlines
Tell-Me
Fig. 1.6 Mean F 1 center frequency in neutral, Tell-Me, and AA sessions (accompanied by
standard deviations in error plots)
also F2 and F3 increase in both genders while remaining relatively steady in
Tell-Me. Note that F1 and F2 increases have been previously reported for stressed
speech, including angry, loud, and Lombard speech modes [ 5 , 14 , 15 ]. Finally,
spectral slopes of the voiced speech segments were extracted by fitting a straight
line to the short-term power spectra in the log amplitude/log frequency plane by
means of linear regression [ 14 ]. The mean spectral slope reaches values around
10.4 dB/Oct, displaying no significant differences across stress classes. Note that
the average slope is somewhat higher than that reported in the literature for clean
neutral speech, presumably due to the strong presence of background car noise, which
introduces additional spectral tilt.
The analysis conducted in this section revealed differences in fundamental
frequency, F 1 ,andF 2 center frequencies between the selected neutral and stressed
classes, confirming that the initial hypothesis about the presence of stress in Tell-Me
and AA segments due to increased cognitive load is valid.
1.3.2 Automatic Classification of Stress
In this section, speech-based neutral/stress classification is proposed and evaluated.
For the purposes of classifier training and testing, the data from 15 drivers were split
into a training set comprising of speech samples from two male and two female
drivers, and test set comprising six male drivers and five female drivers.
Gaussian Mixture Models (GMMs) are chosen to represent probability density
functions (PDFs) of the neutral and stressed classes. The probability of observation
vector o t being generated by the j th GMM is calculated as
M
c jm
ð
T
S 1
jm
e
1
2 ð o t m jm Þ
ð o t m jm Þ
b j ð
o t Þ¼
p
;
(1.2)
Search WWH ::




Custom Search