The effects of training on intelligibility of synthetic speech: I. Pre-test and post-test data

Abstract

Several previous experiments have suggested that the intelligibility of synthetic speech can be improved with practice. However, an alternative interpretation of this research is that subjects simply learned to perform the experimental tasks better without any change in intelligibility. To test these alternatives, we conducted an experiment to separate the effects of training on task performance from improvements in the intelligibility of synthetic speech. Three groups of subjects were tested on day 1 (pre-test) and day 10 (post-test) of the experiment with synthetic speech generated by the Votrax Type-N-Talk text-to-speech system. One group received training with synthetic speech on days 2–9; a second group received exactly the same training procedures on days 2–9 with natural speech; the third group received no training at all. Intelligibility was assessed for isolated words, syntactically correct meaningful sentences, syntactically correct but semantically anomalous sentences, and prose passages. In this paper, we will discuss the effects of training measured by changes in intelligibility between the pre-test and the post-test. The implications of these results for applications of low-cost text-to-speech systems will be discussed.