Audio Similarity Model And Retrieval Based On Emotion

With the development of digital and multimedia technology, multimedia resource hasbecome an important part of the information resource, and music emotion recongnition isan important research field of information retrieval.Most research on music emotion is limited to assign a song with only one emotioncategory, but the emotion expressed by music may be changed as time went on. It will bemore precise if a song corresponds to an emotion sequence.Automatic classification is the basis of music emotion retrieval. We use kinds offeature selection algorithms and classification algorithms on the music dataset, andcompare the results of each combination and finally find out the feature selectionalgorithm and classification algorithm which product highest accuracy. A song, first besegmented by one second and then classify each segment to abtain the initial emotionsequence and then use the smooth algorithm to abtain the finally emotion sequence.We use an emotion sequence similarity model to calculate the music emotionsequence similarity based on the longest common sequence length and the edit distancealgorithms, and we also propose the calculation formula and a series of parameters whichcan be adjusted. In consideration of the similarity and discriminate between each musicemotion category, an convertion cost matix has been added to the edit distance algorithm.Before calculate the accurate similarity, we use two level of filters to filter most of thedissimilarity music to make it faster when retrieval similarity music. Finally, it is provedby experiment that the music emotion sequence similarity model and music similarityretrieval strategy is feasible and effective.