Speech Synthesis Evaluation – Realizing a Social Turn

摘要

Based on a meta-analysis of the state-of-the-art in speech synthesis evaluations, we diagnose the following dilemma: Despite known drawbacks, evaluations predominantly rely on small-scale laboratory tests, typically capturing MOSbased global impressions based on isolated sentences, with the (resynthesized) human voice serving as a gold standard. The problem with such approaches is that synthesis quality can only reliably be estimated if presented in a contextualized manner, e.g. as part of an application and together with its embodiment as an artificial agent, robot or an disembodied voice. As most evaluations are carried out in parallel to system development, and as these tend to be concerned with small details in the developmental process, evaluations of fully fledged applications are often neither possible nor useful. We argue, that with a few modifications in standard evaluation protocols, i.e. by introducing simple interactive scenarios and by relying on both subjective impressionistic and behavioral or physiological measurements, the reliability of such evaluations could be significantly improved.