Towards Multimodal Driver’s Stress Detection - Digital Signal Processing for In-Vehicle Systems and Safety - page 12

Table 1.3 Classification performance; normalized accuracy (%)
Front-end

Expolog
LPC

Expolog
DCT

20Bands
LPC

20Bands
DCT

20Bands
DCT11 NS MFCC PLP

None 83.7 83.1 81.4 81.9 84.2 84.1 83.6

FWSS 85.6 85.1 86.2 85.4 83.5 87.6 88.2

89.0

None
FWSS

86.0

83.0

80.0

77.0

Front-End

Fig. 1.7 Front-end's classification performance

and cepstra extracted from a uniform filterbank of 20 non-overlapping rectangular
filters distributed on a linear frequency scale (20Bands) [ 15 ] were compared. MFCC
represent a common baseline front-end in speech/speaker recognition, and PLP has
been shown by numerous studies to provide comparable or better performance to
MFCC in various speech-related tasks [ 14 ].
Expolog is an outcome of studies on accent classification and stressed speech
recognition, and features based on 20Bands filterbank have shown superior
properties in noisy neutral and Lombard speech recognition [ 15 ].
In this study, Expolog and 20Bands filterbanks were used either as a replacement
for the triangular Mel filterbank in MFCC, yielding front-ends denoted Expolog
DCT and 20Bands DCT, or as a replacement for PLP trapezoid Bark filterbank,
yielding setups denoted Expolog LPC and 20Bands LPC. In order to reduce the
impact of strong background noise on classification, Full Wave Spectral Subtraction
(FWSS) utilizing Burg's cepstral-based voice activity detector [ 14 ] was incorporated
in the feature extraction. The classification results are summarized in Table 1.3 and
Fig. 1.7 . The first row of results in Table 1.3 represents the performance of a classifier
without noise subtraction (NS), denoted “none.”
It can be seen that in the majority of cases, FWSS considerably improves
performance. Among front-ends employing 13 static coefficients and their first-and