Audio content feature selection and classification a random forests and decision tree approach

Content information can be extracted from soundtracks of multimedia files. A good audio classifier as a pre-processor is crucial in such applications. Efforts have been made to develop effective and efficient audio content classifiers in which features were often selected in ad hoc or empirical ways. This paper proposes a set of systematic methods that use the random forests and decision trees to select features and support decisions. The proposed methods allow for heuristic formation of feature spaces, mitigating redundancy in datasets. The performance of the proposed methods has been compared with other common audio classifiers, and improvements in performance have been noted: feature spaces simplified, computational overhead reduced, and classification accuracy improved.