Applying A New Classifier Fusion Technique To Audio Segmentation

This paper presents a new classifier fusion algorithm based on the confusion matrixes of the classifiers which are used to extract the corresponding precision and recall values. The only data needed to be able to apply this new fusion method are the classes or labels assigned by each of the classifiers as well as the reference classes in the development part of the database. The proposed algorithm is described and it is applied to the fusion of two audio segmentation systems that took part in Albayzin 2012 evaluation campaign. The robustness of the algorithm has been assessed and a relative improvement of 6.28% has been achieved when combining the results of the best and worst systems presented to the evaluation.