15th International Congress of Phonetic Sciences (ICPhS-15)

Barcelona, Spain
August 3-9, 2003

WFST Based Unit Selection for Concatenative Speech Synthesis in European Portuguese

Pedro Carvalho, Isabel Trancoso, Luis Oliveira

INESC-ID/IST, Portugal

The goal of the current work is to use Weighted Finite-State Transducers (WFSTs) to model the unit selection task, in a concatenative Text-to-Speech system. One of the major difficulties is the design of a perceptually meaningful cost function that weights and combines several features of the available inventory units, matching them to the target information. The WFST approach allows for great flexibility, as well as an elegant formulation for the unit selection problem. Although there is a price to pay in terms of processing power and memory, one of its main advantages is that it allowed us to experiment with different approaches to the problem.

Full Paper

Bibliographic reference.  Carvalho, Pedro / Trancoso, Isabel / Oliveira, Luis (2003): "WFST based unit selection for concatenative speech synthesis in European Portuguese", In ICPhS-15, 2333-2336.