Audio events classification using hierarchical structure

This paper presents a novel approach using hierarchical structure with different feature sets to classify the audio signals efficiently into audio events. Most of the past researches focused on the speech (male, female), music (different genres) and environment sound (noise). To further differentiate the environment sound, this work studies the feature selection and classification. Different from that of other methods, the audio signals in our work are segmented into different lengths perceptually. So, humans can also recognize a sound based on the short segments. A top-down tree structure with the selected features is designed to classify the audio events while each node is a support vector machine trained as a classifier. Experiments show the robustness and efficiency of the method with a small set of training database.