HYBRID HMM/ANN SYSTEM FOR
OF MACEDONIAN LANGUAGE
Ivan Kraljevski, Dragan Mihajlov, Dejan Gorgjevik
Abstract: This work presents development of a system for artificial speech recognition of Macedonian language. This system uses small vocabulary, speaker dependent and for isolated speech, limited on recognition of digits. The system is based on hybrid architecture combining the Hidden Markov Model with Artificial Neural Networks in order to exploit its advantages. The system transforms the digitalized speech signal into parameters thus obtaining a sequence of acoustical vectors containing information about spectral characteristics. Acoustical vectors are input of the neural network probability classifier. With Dynamical Programming methods the systems chooses the most probable phonetic categories sequence. Then the system using specific criterion chooses a word from the vocabulary which best mach the phonetic sequence. This system can be used in bank-automates for remote financial transactions, military communications, security systems, mathematical applications using voice-input etc.
Keywords: artificial speech recognition, neural networks, hidden Markov model