PROBLEM TO BE SOLVED: To provide a speech recognition accuracy estimating device capable of calculating recognition accuracy as a detailed numerical value after estimating the number of correct answers/substitution errors/insertion errors/deletion errors.SOLUTION: The speech recognition accuracy estimating device comprises: a speech recognition part for creating a word confusion network representing probability that any recognition result word exists for each segment from an inputted speech and probability (existence probability of ε) that any recognition result word does not exist; a word alignment network acquisition part for acquiring a word alignment network representing existence probability of a word with maximum existence probability as correct answer probability, total existence probability of words without maximum existence probability except ε as substitution error probability and the existence probability of ε as insertion error probability when ε has not the maximum existence probability in any segment, and representing total existence probability of the words without maximum existence probability as deletion error probability when ε has the maximum existence probability in any segment; and a random recognition accuracy calculating part for calculating speech recognition accuracy.