Evaluation of classification techniques for audio indexing

This work compares two classification techniques used in audio indexing tasks: Gaussian Mixture Models (GMM) and Support Vector Machines (SVM). GMM is a classical technique taken as reference for comparing the performance of SVM in terms of accuracy and execution time. For testing the methodologies, we perform speech and music discrimination in radio programs and environment sounds (laughter and applause) are identified in TV broadcasts. The objective of the study is to establish references and limits to be considered in practical implementations of audio indexing platforms. Tests show complementary properties between methods and data-driven solutions are suggested as conclusion.