Handling the stochastic behaviour of the Bag-of-Audio-Words method

Automatic sleepiness detection may be of help for drivers and air pilots. Sleepiness can be detected from human speech which is a task of paralinguistics. There are several methods available that can overcome this, like the Bag-of-Audio-Words technique. Because it is a stochastic method it can produce different results after each re-run with the same settings. This can be handled with ensembling. Unfortunately the shared disadvantages of BoAW and ensembling, that BoAW needs quite a large feature space to work well and emsembling makes it even larger and larger which, in the end might lead to a lower classification performance. We found that it can be handled by PCA dimension reduction, but there are some edge cases where it does not work well.