Phonetically conditioned prosody transplantation for TTS: Unit granularity, context and prosody styles

Recently, a `phonetically conditioned' prosody-selection was proposed to realize a target prosody in a conventional TTS output unit-sequence. Here, target prosody-profiles were identified conditioned on their underlying phonetic content with varying degrees of matching and over variable length time-scales that are long as possible. In this paper, we extend these results in two ways: a) we consider various unit-definitions in the 2nd-stage unit-selection with respect to their relative influence on the phonetic conditioning, and hence the nature of prosody selected and, b) show the applicability of the proposed framework to a class of prosodic-styles. The various unit definitions include different granularity (such as fine-category phones to various broad-category phones) and various contextual effects, varying from no context to immediate left-/right-contexts (triphones) and quad-phone contexts. The performance of the proposed framework is characterized using various objective measures, and bring out the effect of the unit-definitions on the prosody selection, and the viability of the framework for a range of prosody styles.