| 摘要 |
PURPOSE: A method for extracting learning data of a voice recognition system is provided to apply a numerical formula and an algorithm for optimizing learning data sets, to study a voice recognizer with fewer learning data sets, so as to uniformly distribute extracted learning data sets to pattern spaces to be studied. CONSTITUTION: A voice recognition system selects predetermined text data from inputted text data to separate the text data, and arranges the text data to set candidate data sets(101, 103). The voice recognition system removes repeated words for words consisting of the candidate data sets, and decides predetermined weight for each word(105). The voice recognition system counts word components of each word having the decided weight, and counts total word components of each word to decide a probability of showing an optional word component(107). The voice recognition system re-selects predetermined words from the candidate data sets, and decides the re-selected words as initial learning data sets(109). The voice recognition system decides gains for the candidate data sets by using the weight and the probability(1111). The voice recognition system repeats the steps "109" and "1111" to decide a learning data set having a maximum gain.
|
