APPARATUS FOR CREATING OPTIMUM ACOUSTIC MODEL BASED ON MAXIMUM LOG LIKELIHOOD AND METHOD THEREOF

摘要

PURPOSE: An optimal acoustic model generating apparatus of a maximum log likelihood base and a method thereof are provided to reduce a calculation during a performance of a voice recognition by estimating 'an acoustic model which has a Gaussian mixture of optimal number' about each tri-phone unit based on a maximum log likelihood estimation. CONSTITUTION: A state occupation probability calculator(104) calculates an occupation probability of each tri-phone unit state. A state log likelihood calculator(106) calculates a log likelihood of each tri-phone unit state. A decision unit of a ML(Maximum Likelihood) mixture by state(110) maps 'the number of mixtures which have the highest log likelihood' in 'the tri-phone unit state which has respective mixture steps'. An optimal acoustic model generator(112) selects a state model of the mapped mixtures from 'the acoustic model of the tri-phone unit state which has various mixture steps' and generates an acoustic model. [Reference numerals] (100) Mixture increasing unit; (102) Repetitive training unit; (104) State occupation probability calculator; (106) State log likelihood calculator; (108) Mixture increase determining unit; (110) State maximum log likelihood mixture determining unit; (112) Optimum sound model generating unit; (AA) Voice data