Measuring the perceptual effects of modelling assumptions in speech synthesis using stimuli constructed from repeated natural speech