Authors:
Dara Pir
1
;
Theodore Brown
2
and
Jarek Krajewski
3
Affiliations:
1
The Graduate Center, City University of New York, United States
;
2
Queens College, City University of New York and The Graduate Center, City University of New York, United States
;
3
University of Wuppertal and Rhenish University of Applied Science Cologne, Germany
Keyword(s):
Automatic Sleepiness Detection, Wrapper Method, Acoustic Group Feature Selection, Computational Paralinguistics.
Abstract:
This paper presents performance results, time complexities, and feature reduction aspects of three wrapper-based acoustic feature selection methods used for automatic sleepiness detection: Between-Groups Feature Selection (BGFS), Within-Groups Feature Selection (WGFS), and Individual Feature Selection (IFS) methods. Furthermore, two different methods are introduced for evaluating system performances. Our systems employ Interspeech 2011 Sleepiness Sub-Challenge’s “Sleepy Language Corpus” (SLC). The two tasks of the wrapper-based method, the feature subset evaluation and the feature space search, are performed by the Support Vector Machine classifier and a fast variant of the Best Incremental Ranked Subset algorithm, respectively. BGFS considers the feature space in Low Level Descriptor (LLD) groups, an acoustically meaningful division, allowing for significant reduction in time complexity of the computationally costly wrapper search cycles. WGFS considers the feature space within each
LLD and generates the feature subset comprised of the best performing individual features among all LLDs. IFS regards the feature space individually. The best classification performance is obtained by BGFS which also achieves improvement over the Sub-Challenge baseline on the SLC test data.
(More)