Croatian Speech Recognition
Ivo Ipšić, Sanda Martinčić-Ipšić · Sciyo eBooks · 2010
Advances in Speech Recognition 124 pronunciation rules sometimes allow more different word accents.Mostly free word order, a complex system of unstressed words and nondeterministic pronunciation rules make the development of pronunciation dictionary and prosodic rules difficult.On the other hand Croatian orthographic rules based on phonological-morphological principle are quite simple which simplifies the definition of orthographic to phonetic rules and process of phonetic transcription.The number of Croatian native speakers is less then 6 millions.Still some interest in the research and development of speech applications for Croatian can be noticed.The speech translation system DIPLOMAT between Serbian and Croatian on one side and English on the other is reported in (Frederking, et al., 1997; Scheytt, et al., 1998;Black, et al., 2002).The TONGUES project continued with this research in direction towards large Croatian vocabulary recognition system.Croatian orthographic-to-phonetic rules are proposed for phonetic dictionary building.The developed Croatian multi-speaker speech corpus was successfully used for the development of speech applications.Proposed Croatian phonetic rules captured adequate Croatian phonetic, linguistic and articulatory knowledge for state tying in acoustical models of the speech recognition system.The Croatian speech recognition system is based on continuous hidden Markov models of context independent (monophones) and context dependent (triphones) acoustic models.The training of speech recognition system was performed using the HTK toolkit (Young et al., 2002; HTK, 2002).Since the main resource in a spoken dialog system design is the collection of speech material, the Croatian speech corpus is presented in Section 2. Orthographic-to-phonetic rules used in the phonetic dictionary preparation are shown as well.Further the acoustic modeling procedures of the speech recognition system including phonetically driven state tying procedures are given in Section 3. Conducted speech recognition experiments and speech recognition results are presented in section 4. We conclude with discussion on advantages of the proposed acoustical modelling approach for Croatian speech recognition and description of current activities and future research plans.