The aim of this thesis is to investigate the possibility of synthesizing speech in the voice of a speaker in a language which he had never spoken. Multilanguage models are created, both for the languages whose databases are annotated using the same conventions, and for the languages whose databases are annotated using different conventions, which includes the Serbian language. Regarding quality of synthesized speech, some models even surpass the quality of synthesis produced by standard monolanguage models. Beside architecture for multilanguage models, Š° method for adaptation of such models to the data of a new speaker is proposed. The proposed method of adaptation enables fast and simple production of new voices, while preserving the possibility to synthesize speech in any language supported by the model, regardless of the target speakerā€™s original language.