Instrumental Quality Estimation for Synthesized Speech Signals

The present thesis summarizes research contributions on objectifying synthesized-speech perception. Based on auditorily derived quality dimension, it is shown how subjective quality rating of synthetic speech can be estimated instrumentally. It turns out that the perceptual peculiarities of synthesized speech, e.g., cognitive nonlinearities, require holistic approaches for deriving the prediction models. Therefore, it is investigated which speech-signal properties best capture the speech quality, and in which way the derived parameters can be beneficially integrated into a non-intrusive prediction model.