Automation Support using non-Invasive Measures of Operator Vocalization (ASIMOV)

Elkin-Frankston, S.1, Leather, C.1, Bracken, B.1, and van Mersbergen, M.2

To be presented at the 11th International Conference on Voice Physiology and Biomechanics (ICVPB 2018), East Lansing, MI (July 2018).

Heightened states of arousal can be investigated using electroglottography (EGG) contact quotient, as well as other measures of vocalization, such as surface laryngeal electromyography (sEMG), and accelerometry (measuring vocal cord vibrations). This type of data can be sensed using voice measures, however sophisticated conditional data processing and interpretation is required to extract these indicators. Current signal processing algorithms developed for laryngeal sEMG data are focused on clinical applications (e.g., diagnosing dysphagia) rather than detecting subtler psychological changes, and cannot readily be adapted to psychologically-based physiological changes. Novel methods are needed to clean and filter data to identify features that support assessment of arousal states. We developed and evaluated models to classify affect using a data from prior studies (van Mersbergen & Delany, 2014), consisting of EGG data, which reflects intrinsic laryngeal musculature and contains information about glottic contact area, and audio data recorded from a microphone. We chose a support vector machine (SVM) to classify data and determine affective state. Preliminary results demonstrate that our models were capable of accurately classifying negative affect. However, while negative affect could be differentiated from neutral affect, classification among positive and negative affect was less accurate. This is likely due to the fact that affect, negative or positive, are both activated through the autonomic nervous system with similar presentations. However, the goal of developing SVM models was to demonstrate our ability to successfully distinguish state. Given the accuracy of our models using limited data sets, we anticipate significant applications and performance gains when applied to new and existing voice data. Results demonstrated the utility of applying machine learning approaches to the field of voice research.

1 Charles River Analytics
2 University of Memphis

For More Information

To learn more or request a copy of a paper (if available), contact Seth Elkin-Frankston.

(Please include your name, address, organization, and the paper reference. Requests without this information will not be honored.)