Audio Retrieval Based on Manifold Ranking

This paper proposes an audio information retrieval model based on Manifold Ranking (MR) and improving ranking results by relevance feedback algorithm. Timbre component has been employed as the main feature. To compute the timbre similarity, it is necessary to extract the spectrum features for each frame. The large set of frames is clustered by a Gaussian Mixture Model (GMM) and Expectation Maximization. The typical spectra frame from GMM is drawn as the data points, manifold ranking assigns each data point a relative ranking score, which is treated as a distance instead of traditional similarity metrics based on pair-wise distance. Furthermore, manifold ranking algorithm can be easily generalized by adding these positive examples by relevance feedback algorithm, and improves the final result. Experimental results show the proposed approach is effective to improve the ranking capability of the existing distance functions.