A speech recognition apparatus is provided with a word dictionary managing subword sequences of words that have possibility of being input together with information indicating importance of each of subwords in each subword sequence which forms a word, where the importance indicates whether or not the subwords of the word are an important factor when recognizing the word, a managing part for managing standard values of acoustic feature patterns of each of the subwords, a generating part for successively reading the words stored in the word dictionary and for generating acoustic feature patterns of each of read subwords using the standard values managed by the managing part, and a collating part for calculating and evaluating distances between an acoustic feature pattern of an input word that is to be recognized and each of the acoustic feature patterns generated by the generating part by referring to the information.