An Experimental Study of Semi-Supervised EM algorithms in Audio Classification and Speaker Identification

Most pattern recognition techniques assume the existence of large quantities of carefully labeled data for training classifiers. However, the generation of his labeled data is an expensive and timet -amounts of data are generated daily, and labeling this data to refine classifiers becomes impossible. In the last years, a new body of techniques has emerged that explore how to take advantage of vast quantities of unlabeled data, i.e. data with no class assignment information. In this paper we study the applicability of these techniques to various audio classification tasks. We show very promising results that demonstrate a reduction in half of audio classification and speaker identification error rates.