In speech recognition systems based on statistical methods is possible to transform the speech recognition problem into the finding the most probable path through a speech recognition network (SRN). The path through the SRN identifies sequence of recognized words. For the construction of this SRN the finite-state transducers (FST) and finite-state automata (FSA), known also as finite-state acceptors can be used. The SRN is constructed before or through speech recognition process from acoustic model, language model, and pronunciation lexicon. In this paper we will describe construction of SRN before speech recognition process based on FST for Slovak language using 100k words pronunciation lexicon. In the end of this paper the resulting SRN and result of testing the SRN will be investigated.
展开▼
