The Blizzard Challenge 2006 CMU Entry introducing hybrid trajectory-selection synthesis
John Kominek, Alan W. Black · 2006
Acknowledging the lessons of Blizzard Challenge 2005 – that smooth prosodic cadence supersedes spectral resolution – but wanting a system devoid of vocoding artifacts – we introduce a hybrid trajectory-selection synthesizer. Using a parametric synthesizer to generate a pitch-synchronous sequence of F0/duration/power and spectral vectors, this trajectory serves as the target cost function for a unit selection synthesizer. The combination can unify the best attributes of two distinct categories of synthesizers, provided that the feature representation supports both. To this end, we also introduce a new perceptually-weighted harmonic representation of speech that is pitch-synchronous and retains phase information.