Adding Personality to Neutral Speech Synthesis Voices

Abstract

A synthetic voice personifies the system using it. Previous work has shown that using sub-corpora with different voice qualities (e.g. tense and lax) can be used to modify the perceived personality of a voice as well as adding expressive and emotional functionality. In this work we explore the use of LPC source/filter decomposition together with modification of the residual to artificially add voice quality sub-corpora to a voice without recording bespoke data. We evaluate this artificially enhanced voice against a baseline unit selection voice with pre-recorded sub-corpora. Although artificial modification impacts naturalness, it has the advantage of adding emotional range to voices where none was recorded in the source data, deals with data sparsity issues caused by sub-corpora, and results in significant effects in terms of perceived emotion.