TALKING FOREIGN - concatenative speech synthesis and the language barrier

This paper presents some solutions to the problem of synthesising multi-lingual speech using waveformconcatenation speech synthesis. The paper presents novel methods for deriving appropriate pronunciations for foreign words in a predominantly native-language text by use of multi-speaker synthesis, and methods for mapping the pronunciations of a foreign-language speaker onto the sounds available in the speech corpus of a native speaker so that the resulting synthesis produces speech which accurately represents the foreign words. The methods differ depending on the language-pair and on the direction of the mapping, because in the case of one-to-many phonemic mappings, high-level features can be used, but in the many-to-one case, a physical representation of the speech signal is required so that use can be made of the natural variability in speech production to select the most appropriate allophonic variant. All mappings are automatic, and the use of rule-based procedures which require human knowledge is minimised. In this way, the methods are extensible to any language combinations. Synthesised speech samples are included with the paper so that subjective evaluation of the results can be made.