Dominant Feature Vectors Based Audio Similarity Measure

This paper presents an approach to extracting dominant feature vectors from an individual audio clip and then proposes a new similarity measure based on the dominant feature vectors. Instead of using the mean and standard deviation of frame features in most conventional methods, the most salient characteristics of an audio clip are represented in the form of several dominant feature vectors. These dominant feature vectors give a better description of the fundamental properties of an audio clip, especially when frame features change a lot along the time line. Evaluations on a content-based audio retrieval system indicate an obvious improvement after using the proposed similarity measure, compared with some other conventional methods.