Monitoring eye movements as an evaluation of synthesized speech

We describe a novel empirical method for synthesized speech evaluation that relies on monitoring participant eye movements as they respond to spoken instructions in a visual workspace. We use instruction sets generated with two different text-to-speech synthesizers and a human voice for baseline comparison. We replicate findings demonstrating incremental processing for natural speech and at the same time demonstrate incremental processing for synthesized speech for both lexical and discourse-level processing. Results reveal differences in the time course of processing between natural and synthesized voices as well as between the two synthesized voices. These results, including the detection of subtle processing differences between voices, demonstrate the potential for eye movement monitoring as a promising new methodology that provides an on-line measure of synthesized speech processing more fine-grained than possible with techniques used to date.