Influence of Features Extraction Methods in Performance of Continuous Speech Recognition for Romanian

This paper describes continuous speech recognition experiments for Romanian language, based on statistical modelling by using hidden Markov models. These experiments are made in order to select the most appropriate features extraction method. The compared methods are cepstral and LPC analysis, in standard and perceptual versions. In our tests the cepstral coefficients perform in the most situations better versus the linear prediction ones, and the perceptual coefficients better than the standard ones.