Submission to Mirex Ams Task 2011 – Sparse Coding Similarity Learning Method

This paper describes our submissions to the MIREX 2011 audio music similarity and retrieval task. The proposed method is based on a machine learning technique – sparse coding (SC). The music similarity is not directly obtained from computed distance measures on audio contents, instead, we predict a higher level similarity scores to match the listener’s subjective perceptions based on these distance measures and our pre-trained models. At the training stage, we will record the mapping of computed distance measures and the associated high level similarity scores which is estimated from human (expert) tags. Using the sparse coding techniques, this mapping is recorded using two jointly learned dictionaries that respectively store the representative computed distance and the high level similarity score patterns. Both the computed distance measures and the similarity scores can be represented as a sparse linear combination of the elements in the corresponding dictionary. At the testing stage, we will find the ratios that each element in the dictionary contributes to the newly computed distance measures, and then use this ratio to predict the corresponding similarity scores. 1. ALGORITHM OVERVIEW We illustrate the flowchart of our algorithm in Figure 1. At the training stage, we first take the expert labelled tags of the training audio to estimate multifaceted similarity scores, such as, general-similarity, style-similarity, moodsimilarity, etc. Then, for each song, we extract a set of computable acoustic features. Using these features, we compute multiple distance measures for each song-pair. Next, we jointly learn two dictionaries for the representative distance measure patterns and the associated similarity score patterns. Given any two testing songs in the testing stage, we will compute its distance measures to estimate the similarity scores. This document is licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 License. http://creativecommons.org/licenses/by-nc-sa/3.0/ c © 2010 The Authors. Acoustic