Polyglot Voice Design for Unit Selection Speech Synthesis

Text-to-speech (TTS) systems are increasingly faced with mixed language input. Most TTS systems allow building synthetic voices for different languages, but each voice \"speaks\" only one language at a time. For synthesizing mixed language input polyglot voices are needed that are able to switch between languages.An inherent problem of unit selection speech synthesis is that its quality depends on the contents of the unit database. Concatenation of units not in the database usually results in bad synthesis quality. However, the size of a database with good unit coverage is prohibitively large if the intended domain is unrestricted. In polyglot databases not only single language units have to be stored, but also the concatenation points of words from different languages. Therefore it is worth exploring if database size can be reduced by including only single language units and handling multilingual units at synthesis time.This work examines alternative methods for handling multilingual cross-word units. Three approaches are proposed and evaluated in perceptual experiments. Results suggest that proposed methods can replace full coverage of units in unit selection databases.