Interspeech'2005 - Eurospeech
In this paper we describe the role of multilingual models in the creation and deployment of unit selection based bilingual speech synthesizers. We first review the definition of a multilingual phonetic alphabet for the simultaneous recognition of up to fifteen languages, and then discuss synthesis specific modifications that allow a more detailed description of the synthesizers' unit inventories. Furthermore, we investigate on the use of multilingual phonological decision trees for a fast identification of candidate segments, and for the prediction of pitch and duration targets during synthesis. Experimental results for two different combinations of languages demonstrate the feasibility of multilingual modeling for speech synthesis.
Bibliographic reference. Ordinas, J. Botella / Fischer, V. / Waast-Richard, C. (2005): "Multilingual models in the IBM bilingual text-to-speech systems", In INTERSPEECH-2005, 1485-1488.