|
KEY WORD RECOGNITION IN THE CONTINUAL SPEECH IN SERBIAN LANGUAGE
This paper presents the latest results obtained in the realization of an algorithm for the recognition of certain number of key words in the continual speech in Serbian language. The essence of the applied procedure is to model key words in an uttered sentence by the hidden Markov models (HMMs) and to calculate reliability for the whole sentence in the decoding phase. By comparing the reliability measure with the treshold, determined in the training phase, it can be determined whether one of key words is pronounced in an uttered sentence. Key words are modelled by the continual HMMs which are obtained by relating HMMs with three states for each syllable a key word is made of. The measure of system’s quality (MSQ) is introduced in order to determine the optimal step and treshold for the reliability measure in the decoding phase.
Ključne reči: Key word recognition, recognition algorithm, hidden Markov model, Serbian language, man-machine dialogue. |