Short description

One reason Electro-Larynx (EL) speech sound robotic is the lack of a natural pitch contour, because commercially available EL devices have only limited pitch variation.
Previous research on pitch perception in whispered speech has shown that even though no harmonic excitation is present, something like a pitch sensation can be documented. This approach uses this information and derives a fundmental frequency (F0) contour from speech features, which are not related to the fundamental frequency, but substitute the F0 in conveying prosodic information. This generated F0 contour is then used to drive the EL excitation. The student should choose useful speech features, which are mapped to the F0 contour by using a machine learning approach.

Your Tasks

• Review of literature
• Implementation of an approach using Matlab
• Evaluation on a provided database

Your Profile/Prerequisites

• Speech Communication 1
• Speech Communication 2
• Speech Communication Laboratory is recommended
• Matlab

Contact:

Martin Hagmüller, Franz Pernkopf ({hagmueller,pernkopf}@tugraz.at or 0316/873 4377)