From Low-Level to High-Level: Comparative Study of Music Similarity Measures

Abstract

Studying the ways to recommend music to a user is a central task within the music information research community. From a content-based point of view, this task can be regarded as obtaining a suitable distance measurement between songs defined on a certain feature space. We propose two such distance measures. First, a low-level measure based on tempo-related aspects, and second, a highlevel semantic measure based on regression by support vector machines of different groups of musical dimensions such as genre and culture, moods and instruments, or rhythm and tempo. We evaluate these distance measures against a number of state-of-the-art measures objectively, based on 17 ground truth musical collections, and subjectively, based on 12 listeners' ratings. Results show that, in spite of being conceptually different, the proposed methods achieve comparable or even higher performance than the considered baseline approaches. Furthermore, they open up the possibility to explore distance metrics that are based on truly semantic notions.

Semantic audio content-based music recommendation and visualization based on user preference examples

“…Second, we assume that high-level semantic description outperforms common low-level feature information in the task of music recommendation. The latter hypothesis is based on similar evidence in the case of music similarity estimation (Bogdanov et al, 2009).…”

Unifying Low-Level and High-Level Music Similarity Measures

“…Altogether, the classification results form a high-level descriptor space, which contains the probability values of each class for each SVM. Based on results in [35], we decided to use the libSVM 7 implementation with the C-SVC method and a radial basis function kernel with default parameters.…”

“…We evaluated all considered approaches with a uniform methodological basis, including an objective evaluation on comprehensive ground truths and a subjective evaluation based 7 http://www.csie.ntu.edu.tw/ ∼ cjlin/libsvm/ on ratings given by real listeners. As an initial benchmark for the comparison of the considered approaches we used a random distance (RAND), i.e.…”

“…In the second part, which we partially presented in [7], we compare the aforementioned baseline approaches against two novel distance measures (Sec. III).…”

A content-based system for music recommendation and visualization of user preferences working on semantic notions

“…We follow [2] to search for the tracks from our in-house music collections which are similar to the user's preference set. We use a semantic distance, proposed in [3,4] and implemented in the core of the similarity engine provided by the Canoris API. This similarity engine supports queries-by-example to retrieve the most similar files based on the analysis results.…”