This paper describes the Text-to-Speech (TTS) systems presented by the Buceador Consortium in the Blizzard Challenge 2011 evaluation campaign. The main system is a concatenative hybrid one that tries to combine the strong points of both statistical and unit selection synthesis (i.e. robustness and segmental naturalness respectively). The hybrid system has reached results significantly above average as far as similarity and naturalness are concerned, with no significant differences with most of the systems in the intelligibility task. This clearly improves the performance achieved in previous participations, and shows the validity of the hybrid approach proposed. Besides, an HMM-based system was built for the ES1 intelligibility tasks, using an HNM-based vocoder.

Peer Reviewed