A large-vocabulary speech recognition method on the basis of a multi-layer central lexicon and an apparatus therefor are provided to perform detailed matching only for vocabularies of a region similar to input utterance through a symbol matching process of a phoneme recognized result and a central lexicon in a tree structure. A multi-pass symbol matching unit(140) performs the multi-pass symbol matching between a phoneme recognition string and a pronunciation string of a central lexicon layered as a tree structure. A detailed matching unit(150) performs detail matching for selecting a speed recognition result by using a detailed acoustic model for candidate vocabulary sets selected by the multi-pass symbol matching. The multi-pass symbol matching unit(140) tracks a node which records the maximum matching score per each layer, and repeats a node tracking process until the multi-pass symbol matching unit(140) arrives at the terminating node.
展开▼