Date of Degree


Document Type


Degree Name



Computer Science


Theodore Brown

Committee Members

Victor Pan

Xiaowen Zhang

Wei Dong

Subject Categories

Computer Sciences


feature selection, computational paralinguistics


The burgeoning field of computational paralinguistics deals with the ways in which spoken words are uttered and attempts to recognize the states and traits of the speakers. Many areas of current scientific research, including computational paralinguistics, have started to employ datasets with ever increasing number of features. Using large feature sets has helped improve recognition performances. However, processing these large sets has given rise to various problems. Feature selection methods, which reduce the dimensionality of the original feature sets by removing irrelevant and/or redundant features, could be used to address these problems.

The two main methods for feature selection are the wrapper and the filter methods. The wrapper method uses accuracy predictions that a classifier generates in order to evaluate feature subsets, whereas the filter method uses characteristics of the data alone. Wrappers can provide superior classification performances as they use the biases of a learning algorithm but are operationally costly especially when combined with computationally complex classifiers. Filters do not require the aid of a learning algorithm and can therefore operate at faster speeds.

We propose wrapper-based methods that consider the feature space in acoustically meaningful groups rather than individually and render the computationally intensive classification process more tractable. In addition, we demonstrate other advantages of the group-based feature selection approach as compared to those using individual features. Furthermore, we propose variants of the group-based method for classification in the presence of noise and compare their performances with alternative filter methods.

This work is embargoed and will be available for download on Sunday, June 02, 2019

Graduate Center users:
To read this work, log in to your GC ILL account and place a thesis request.

Non-GC Users:
See the GC’s lending policies to learn more.