Learning the Similarity of Audio Music in Bag-of-frames Representation from Tagged Music Data.

Due to the cold-start problem, measuring the similarity between two pieces of audio music based on their low-level acoustic features is critical to many Music Information Retrieval (MIR) systems. In this paper, we apply the bag-offrames (BOF) approach to represent low-level acoustic features of a song and exploit music tags to help improve the performance of the audio-based music similarity computation. We first introduce a Gaussian mixture model (GMM) as the encoding reference for BOF modeling, then we propose a novel learning algorithm to minimize the similarity gap between low-level acoustic features and music tags with respect to the prior weights of the pre-trained GMM. The results of audio-based query-by-example MIR experiments on the MajorMiner and Magnatagatune datasets demonstrate the effectiveness of the proposed method, which gives a potential to guide MIR systems that employ BOF modeling.