An improvement of speech characteristics of the Croatian speech synthesis engine

Strojna tvorba govora je proces umjetne proizvodnje govora. Pokušaji izrade sustava za strojnu tvorbu govora sežu u doba prije računala, međutim, danas se taj naziv prvenstveno odnosi na računalnu tvorbu govora. Javni sustav za strojnu tvorbu hrvatskoga govora HascheckVoice već je nekoliko godina u upotrebi. Premda sustav sintetizira razumljivi govor, njegove govorne karakteristike daleko su od poželjnih. U sklopu ovoga rada pokušale su se popraviti njegove karakteristike uz pomoć novokreiranog govornog uzorka u trajanju od preko jednog sata govora treniranog govonika. Tijekom izrade rješenja proučena su i istražena naprednije mogućnosti Festivala i Festvox, te su zapažanja navedena u radu, jednako kao i probleme na koje se naišlo i potencijalna rješenja i put za daljnji razvitak glasa.Machine generated speech is the process of artificial speech production. Attempts at making such systems date back to the era before computers. However, now such systems are primarily related to the production of computer speech. Public system for automated Croatian speech synthesis HascheckVoice has already been in use for several years. Although the system produces intelligible synthesized speech, its voice characteristics are far from desirable. The goal of this study was to try to fix its features with the help of newly created voice patterns for a period of over one hour recorded by a trained speaker. Whilst preparing this paper, advanced features of Festival and Festvox have been researched and noted in this paper, as well as the problems encountered and possible solutions