The objective of the present invention is to improve accuracy of unsupervised abnormal sound detection using a small number of items of abnormal sound data. A threshold setting unit (13) uses a normal model that has learned using normal sound data, and an abnormal model representing abnormal sound data to calculate a degree of abnormality for each of a plurality of items of abnormal sound data, and sets the minimum value thereof as a threshold. A weighting updating unit (14) uses a plurality of items of normal sound data, the abnormal sound data and the threshold to update weightings in the abnormal model such that all the abnormal sound data are determined to be abnormal and the probability that normal sound data are determined to be abnormal is minimized.
展开▼