A Comparative Study of Three Speech Recognition Systems for Romanian Language

In recent years, there were different concerns to develop robust automatic speech recognition (ASR) systems for Romanian language, but only a few have been implemented for Romanian language. In this study, we present three ASR systems, implemented for Romanian language and their results are compared. The first one is a study of vowels and digits recognition, using artificial neural networks (ANN), developed by G. Toderean from Technical University of Cluj-Napoca, in collaboration with M. Giurgiu and M. Costeiu. The second one was developed by C. Dumitru from Politehnica Bucuresti University and is a system containing three modules: digits recognition, vowels recognition and telephone numbers recognition. These modules use statistical models - HMM (Hidden Markov Models), neural models ANN and hybrid models (statistical and neural models). The third system was designed by H. N. Teodorescu, from Technical University of Iasi, and contains the recognition of the basic sounds (vowels, consonants and specific sounds), short sentences or parts of sentences with different emotional charge, pathological voices. The aim of this comparison is to find differences and similarities between these systems.