Expressive speech synthesis using American English ToBI: Questions and contrastive emphasis

Abstract

We describe American English concatenative text-to-speech synthesis experiments in which "expressions," here, questioning and contrastive emphasis, are each associated with a ToBI prosodic template. ToBI labels, along with text features, are in turn incorporated into decision-tree models of F0 and segment duration to be used during synthesis, sparing the need for expression-specific large corpora and decision trees. Synthesizing using this approach enables listeners to perform the difficult task of distinguishing yes-no questions from identically-worded declarative sentences 78% of the time, compared to the baseline system's 50%. For contrastive emphasis, a sentence is synthesized with emphasis on a word which is chosen appropriately or inappropriately based on a preceding sentence. Listeners' mean opinion scores for appropriate emphases exceed inappropriate by 0.40 on a 1-to-5 scale for the experimental system, compared to a difference of 0.11 for the baseline, a significant system difference (p < 0.01).