Features Extraction, Modeling and Training Strategies in Continuous Speech Recognition for Romanian Language

This paper describes continuous speech recognition experiments for Romanian language, by using HMM modeling. The following questions are to be discussed: the realization of a new front-end reconsidering linear prediction, the enhancement of recognition rates by context dependent modeling, the evaluation of training strategies ensuring speaker independence of the recognition process without speaker adaptation procedures, by speaker selection for training. The experiments lead to a development of the initial system with a promising front-end based on PLP coefficients, second ranked for the recognition performance obtained, near the first ranked front-end based on mel-frequency cepstral coefficients (MFCC), but far better as the last ranked, based on simple linear prediction. Concerning the implemented algorithm for context dependent modeling, it permits in all situations enhanced recognition rates. The experiments made with gender speaker selection enhanced under certain conditions the recognition rate, proving good generalization properties especially by training with the male speakers database