Abstract:

Extensive lectures influence a speaker’s voice quality after a certain period of time. In this thesis the voice changes of individual speakers should be analyzed by generating and evaluating a suitable feature space in order to enable a voice quality characterization in real-time. First, a convenient audio data set should be recorded (e.g. 10 speaker recordings in a lecture hall) and labeled (subjective and with the help of speech therapists at the MedUni Graz). Secondly, vocal fatigue related features must be found within a literature research (feature extraction) and the extracted parameters should be analyzed with statistical methods to determine a beneficial feature set for a binary classification task (feature selection/ranking). Finally, the data should be classified with the use of state-of-the-art machine learning algorithms in the sense of supervised learning.

Tasks:

• organisation (ca. 10 test subjects) and provision of useful audio recordings (ca. 2h) with the help of a mobile recording system (cooperation with the company audEERING)
• literature research (vocal fatigue, vocal loading, voice dosimetry, voice monitoring, …) in order to extract a suitable feature set (e.g. with openSMILE)
• analyzing the feature space with statistical methods
• binary classification with the help of machine learning algorithms