| Popis: |
Both ASR and TTS systems described in this chapter have been originally developed for the Serbian language. However, linguistic similarities among South Slavic languages have allowed the adaptation of this system to other South Slavic languages, with various degrees of intervention needed. As for ASR, adaptation to Bosnian and Croatian was very simple (due to extreme similarity of phonetics), whereas for Macedonian it was necessary to develop separate speech databases. The actual procedures used for ASR were almost identical in all cases. While well known algorithms were used for model training and testing, in this chapter only the original algorithms are presented. The VTN procedure based on the use of the iterative method and only static features for VTN coefficient estimation shows significant improvement in comparison to the common VTN procedure. The eigenvalue driven Gaussian selection significantly reduce computational load with minor increase of WER. Neither of the proposed algorithms is language dependent. As for TTS, conversion of an arbitrary text into intelligible and natural-sounding speech has proven to be a highly language-dependent task, and the degree of intervention was variable and depended on specific properties of a particular language. For example, the simplicity of accentuation in Macedonian has allowed POS tagging and syntactic parsing to be avoided altogether, at the price of certain impairment in quality of synthesis. On the other hand, for Croatian and Bosnian, it was also necessary to build new accentuation dictionaries and to revise the expert system for POS tagging in order to assign words their appropriate accentuation, necessary for production of natural sounding speech. It can be concluded that, in spite of the apparent language dependence of both principal speech technologies, some of their segments can be developed in parallel or re-used. The ASR and TTS systems described here are widely applied across the Western Balkans. In fact, practically all applications of speech technologies in the countries of the Western Balkans (Pekar et al., 2010) are based on ASR and TTS components described in this chapter. |
