Prosody transplantation using unit-selection: Principles and early results
Mythri Thippareddy, Viswanathan Ramasubramanian · 2015
We address the problem of prosody transplantation or conversion in TTS in a 2-stage unit-selection framework, which uses the units determined in a 1st-stage conventional unit-selection to be mapped to units in a 2nd-stage prosodic unit-selection. This retrieves units from the 2nd-stage unit-database that has an associated prosody in the prosodic-style of the 2nd-stage unit-database and allows further incorporation of the prosodic information from the 2nd-stage unit-selection on to the spectral information of the 1st-stage unit-selection in any of the conventional synthesizer system. We introduce a Viterbi algorithm for optimal unit-selection on pre-grouped units and identify several experimental scenarios that will impact the nature of the prosody that can be transplanted from the 2nd-stage unit-selection. Preliminary results confirm the validity of the framework, and show the promise of such a prosody transplantation scheme.