Audio retrieval: based on unsupervised learning approach

An efficient audio retrieval algorithm through unsupervised learning was proposed. The procedure of this algorithm has the following procedures: firstly, extracting features directly from the compressed domain; secondly, generating limited number of centroids through the clustering of minimum spanning tree (MST), and the clustering centroids are used to represent each audio clip and performed efficient matching of audio clips; finally, in order to guarantee the retrieval results consistent with the user´s subjective perception, the update of feature weights and centroids are performed. Experiments show that the audio retrieval method is robust against noise.